6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciObjArray.hpp"
27 #include "ci/ciUtilities.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "compiler/compileLog.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/c2/barrierSetC2.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "opto/addnode.hpp"
35 #include "opto/castnode.hpp"
36 #include "opto/convertnode.hpp"
37 #include "opto/graphKit.hpp"
38 #include "opto/idealKit.hpp"
39 #include "opto/intrinsicnode.hpp"
40 #include "opto/locknode.hpp"
41 #include "opto/machnode.hpp"
42 #include "opto/opaquenode.hpp"
43 #include "opto/parse.hpp"
44 #include "opto/reachability.hpp"
45 #include "opto/rootnode.hpp"
46 #include "opto/runtime.hpp"
47 #include "opto/subtypenode.hpp"
48 #include "runtime/deoptimization.hpp"
49 #include "runtime/sharedRuntime.hpp"
50 #include "utilities/bitMap.inline.hpp"
51 #include "utilities/growableArray.hpp"
52 #include "utilities/powerOfTwo.hpp"
53
54 //----------------------------GraphKit-----------------------------------------
55 // Main utility constructor.
56 GraphKit::GraphKit(JVMState* jvms)
57 : Phase(Phase::Parser),
58 _env(C->env()),
59 _gvn(*C->initial_gvn()),
60 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
61 {
62 _exceptions = jvms->map()->next_exception();
63 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
64 set_jvms(jvms);
65 }
66
67 // Private constructor for parser.
68 GraphKit::GraphKit()
69 : Phase(Phase::Parser),
70 _env(C->env()),
71 _gvn(*C->initial_gvn()),
72 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
73 {
74 _exceptions = nullptr;
75 set_map(nullptr);
76 DEBUG_ONLY(_sp = -99);
77 DEBUG_ONLY(set_bci(-99));
78 }
79
80
81
82 //---------------------------clean_stack---------------------------------------
83 // Clear away rubbish from the stack area of the JVM state.
84 // This destroys any arguments that may be waiting on the stack.
85 void GraphKit::clean_stack(int from_sp) {
86 SafePointNode* map = this->map();
87 JVMState* jvms = this->jvms();
88 int stk_size = jvms->stk_size();
89 int stkoff = jvms->stkoff();
90 Node* top = this->top();
91 for (int i = from_sp; i < stk_size; i++) {
92 if (map->in(stkoff + i) != top) {
93 map->set_req(stkoff + i, top);
94 }
95 }
96 }
97
98
99 //--------------------------------sync_jvms-----------------------------------
100 // Make sure our current jvms agrees with our parse state.
329 }
330 static inline void add_one_req(Node* dstphi, Node* src) {
331 assert(is_hidden_merge(dstphi), "must be a special merge node");
332 assert(!is_hidden_merge(src), "must not be a special merge node");
333 dstphi->add_req(src);
334 }
335
336 //-----------------------combine_exception_states------------------------------
337 // This helper function combines exception states by building phis on a
338 // specially marked state-merging region. These regions and phis are
339 // untransformed, and can build up gradually. The region is marked by
340 // having a control input of its exception map, rather than null. Such
341 // regions do not appear except in this function, and in use_exception_state.
342 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
343 if (failing_internal()) {
344 return; // dying anyway...
345 }
346 JVMState* ex_jvms = ex_map->_jvms;
347 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
348 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
349 assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
350 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
351 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
352 assert(ex_map->req() == phi_map->req(), "matching maps");
353 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
354 Node* hidden_merge_mark = root();
355 Node* region = phi_map->control();
356 MergeMemNode* phi_mem = phi_map->merged_memory();
357 MergeMemNode* ex_mem = ex_map->merged_memory();
358 if (region->in(0) != hidden_merge_mark) {
359 // The control input is not (yet) a specially-marked region in phi_map.
360 // Make it so, and build some phis.
361 region = new RegionNode(2);
362 _gvn.set_type(region, Type::CONTROL);
363 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
364 region->init_req(1, phi_map->control());
365 phi_map->set_control(region);
366 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
367 record_for_igvn(io_phi);
368 _gvn.set_type(io_phi, Type::ABIO);
369 phi_map->set_i_o(io_phi);
899 if (PrintMiscellaneous && (Verbose || WizardMode)) {
900 tty->print_cr("Zombie local %d: ", local);
901 jvms->dump();
902 }
903 return false;
904 }
905 }
906 }
907 return true;
908 }
909
910 #endif //ASSERT
911
912 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
913 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
914 ciMethod* cur_method = jvms->method();
915 int cur_bci = jvms->bci();
916 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
917 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
918 return Interpreter::bytecode_should_reexecute(code) ||
919 (is_anewarray && code == Bytecodes::_multianewarray);
920 // Reexecute _multianewarray bytecode which was replaced with
921 // sequence of [a]newarray. See Parse::do_multianewarray().
922 //
923 // Note: interpreter should not have it set since this optimization
924 // is limited by dimensions and guarded by flag so in some cases
925 // multianewarray() runtime calls will be generated and
926 // the bytecode should not be reexecutes (stack will not be reset).
927 } else {
928 return false;
929 }
930 }
931
932 // Helper function for adding JVMState and debug information to node
933 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
934 // Add the safepoint edges to the call (or other safepoint).
935
936 // Make sure dead locals are set to top. This
937 // should help register allocation time and cut down on the size
938 // of the deoptimization information.
939 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
1015 uint p = debug_start; // walks forward in [debug_start, debug_end)
1016 uint j, k, l;
1017 SafePointNode* in_map = in_jvms->map();
1018 out_jvms->set_map(call);
1019
1020 if (can_prune_locals) {
1021 assert(in_jvms->method() == out_jvms->method(), "sanity");
1022 // If the current throw can reach an exception handler in this JVMS,
1023 // then we must keep everything live that can reach that handler.
1024 // As a quick and dirty approximation, we look for any handlers at all.
1025 if (in_jvms->method()->has_exception_handlers()) {
1026 can_prune_locals = false;
1027 }
1028 }
1029
1030 // Add the Locals
1031 k = in_jvms->locoff();
1032 l = in_jvms->loc_size();
1033 out_jvms->set_locoff(p);
1034 if (!can_prune_locals) {
1035 for (j = 0; j < l; j++)
1036 call->set_req(p++, in_map->in(k+j));
1037 } else {
1038 p += l; // already set to top above by add_req_batch
1039 }
1040
1041 // Add the Expression Stack
1042 k = in_jvms->stkoff();
1043 l = in_jvms->sp();
1044 out_jvms->set_stkoff(p);
1045 if (!can_prune_locals) {
1046 for (j = 0; j < l; j++)
1047 call->set_req(p++, in_map->in(k+j));
1048 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1049 // Divide stack into {S0,...,S1}, where S0 is set to top.
1050 uint s1 = stack_slots_not_pruned;
1051 stack_slots_not_pruned = 0; // for next iteration
1052 if (s1 > l) s1 = l;
1053 uint s0 = l - s1;
1054 p += s0; // skip the tops preinstalled by add_req_batch
1055 for (j = s0; j < l; j++)
1056 call->set_req(p++, in_map->in(k+j));
1057 } else {
1058 p += l; // already set to top above by add_req_batch
1059 }
1060
1061 // Add the Monitors
1062 k = in_jvms->monoff();
1063 l = in_jvms->mon_size();
1064 out_jvms->set_monoff(p);
1065 for (j = 0; j < l; j++)
1066 call->set_req(p++, in_map->in(k+j));
1067
1255 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1256 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1257 return _gvn.transform( new AndLNode(conv, mask) );
1258 }
1259
1260 Node* GraphKit::ConvL2I(Node* offset) {
1261 // short-circuit a common case
1262 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1263 if (offset_con != (jlong)Type::OffsetBot) {
1264 return intcon((int) offset_con);
1265 }
1266 return _gvn.transform( new ConvL2INode(offset));
1267 }
1268
1269 //-------------------------load_object_klass-----------------------------------
1270 Node* GraphKit::load_object_klass(Node* obj) {
1271 // Special-case a fresh allocation to avoid building nodes:
1272 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1273 if (akls != nullptr) return akls;
1274 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1275 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS));
1276 }
1277
1278 //-------------------------load_array_length-----------------------------------
1279 Node* GraphKit::load_array_length(Node* array) {
1280 // Special-case a fresh allocation to avoid building nodes:
1281 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1282 Node *alen;
1283 if (alloc == nullptr) {
1284 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1285 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1286 } else {
1287 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1288 }
1289 return alen;
1290 }
1291
1292 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1293 const TypeOopPtr* oop_type,
1294 bool replace_length_in_map) {
1295 Node* length = alloc->Ideal_length();
1304 replace_in_map(length, ccast);
1305 }
1306 return ccast;
1307 }
1308 }
1309 return length;
1310 }
1311
1312 //------------------------------do_null_check----------------------------------
1313 // Helper function to do a null pointer check. Returned value is
1314 // the incoming address with null casted away. You are allowed to use the
1315 // not-null value only if you are control dependent on the test.
1316 #ifndef PRODUCT
1317 extern uint explicit_null_checks_inserted,
1318 explicit_null_checks_elided;
1319 #endif
1320 Node* GraphKit::null_check_common(Node* value, BasicType type,
1321 // optional arguments for variations:
1322 bool assert_null,
1323 Node* *null_control,
1324 bool speculative) {
1325 assert(!assert_null || null_control == nullptr, "not both at once");
1326 if (stopped()) return top();
1327 NOT_PRODUCT(explicit_null_checks_inserted++);
1328
1329 // Construct null check
1330 Node *chk = nullptr;
1331 switch(type) {
1332 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1333 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1334 case T_ARRAY : // fall through
1335 type = T_OBJECT; // simplify further tests
1336 case T_OBJECT : {
1337 const Type *t = _gvn.type( value );
1338
1339 const TypeOopPtr* tp = t->isa_oopptr();
1340 if (tp != nullptr && !tp->is_loaded()
1341 // Only for do_null_check, not any of its siblings:
1342 && !assert_null && null_control == nullptr) {
1343 // Usually, any field access or invocation on an unloaded oop type
1344 // will simply fail to link, since the statically linked class is
1345 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1346 // the static class is loaded but the sharper oop type is not.
1347 // Rather than checking for this obscure case in lots of places,
1348 // we simply observe that a null check on an unloaded class
1412 }
1413 Node *oldcontrol = control();
1414 set_control(cfg);
1415 Node *res = cast_not_null(value);
1416 set_control(oldcontrol);
1417 NOT_PRODUCT(explicit_null_checks_elided++);
1418 return res;
1419 }
1420 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1421 if (cfg == nullptr) break; // Quit at region nodes
1422 depth++;
1423 }
1424 }
1425
1426 //-----------
1427 // Branch to failure if null
1428 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1429 Deoptimization::DeoptReason reason;
1430 if (assert_null) {
1431 reason = Deoptimization::reason_null_assert(speculative);
1432 } else if (type == T_OBJECT) {
1433 reason = Deoptimization::reason_null_check(speculative);
1434 } else {
1435 reason = Deoptimization::Reason_div0_check;
1436 }
1437 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1438 // ciMethodData::has_trap_at will return a conservative -1 if any
1439 // must-be-null assertion has failed. This could cause performance
1440 // problems for a method after its first do_null_assert failure.
1441 // Consider using 'Reason_class_check' instead?
1442
1443 // To cause an implicit null check, we set the not-null probability
1444 // to the maximum (PROB_MAX). For an explicit check the probability
1445 // is set to a smaller value.
1446 if (null_control != nullptr || too_many_traps(reason)) {
1447 // probability is less likely
1448 ok_prob = PROB_LIKELY_MAG(3);
1449 } else if (!assert_null &&
1450 (ImplicitNullCheckThreshold > 0) &&
1451 method() != nullptr &&
1452 (method()->method_data()->trap_count(reason)
1486 }
1487
1488 if (assert_null) {
1489 // Cast obj to null on this path.
1490 replace_in_map(value, zerocon(type));
1491 return zerocon(type);
1492 }
1493
1494 // Cast obj to not-null on this path, if there is no null_control.
1495 // (If there is a null_control, a non-null value may come back to haunt us.)
1496 if (type == T_OBJECT) {
1497 Node* cast = cast_not_null(value, false);
1498 if (null_control == nullptr || (*null_control) == top())
1499 replace_in_map(value, cast);
1500 value = cast;
1501 }
1502
1503 return value;
1504 }
1505
1506
1507 //------------------------------cast_not_null----------------------------------
1508 // Cast obj to not-null on this path
1509 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1510 const Type *t = _gvn.type(obj);
1511 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1512 // Object is already not-null?
1513 if( t == t_not_null ) return obj;
1514
1515 Node* cast = new CastPPNode(control(), obj,t_not_null);
1516 cast = _gvn.transform( cast );
1517
1518 // Scan for instances of 'obj' in the current JVM mapping.
1519 // These instances are known to be not-null after the test.
1520 if (do_replace_in_map)
1521 replace_in_map(obj, cast);
1522
1523 return cast; // Return casted value
1524 }
1525
1526 // Sometimes in intrinsics, we implicitly know an object is not null
1527 // (there's no actual null check) so we can cast it to not null. In
1528 // the course of optimizations, the input to the cast can become null.
1529 // In that case that data path will die and we need the control path
1584 Node* GraphKit::memory(uint alias_idx) {
1585 MergeMemNode* mem = merged_memory();
1586 Node* p = mem->memory_at(alias_idx);
1587 assert(p != mem->empty_memory(), "empty");
1588 _gvn.set_type(p, Type::MEMORY); // must be mapped
1589 return p;
1590 }
1591
1592 //-----------------------------reset_memory------------------------------------
1593 Node* GraphKit::reset_memory() {
1594 Node* mem = map()->memory();
1595 // do not use this node for any more parsing!
1596 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1597 return _gvn.transform( mem );
1598 }
1599
1600 //------------------------------set_all_memory---------------------------------
1601 void GraphKit::set_all_memory(Node* newmem) {
1602 Node* mergemem = MergeMemNode::make(newmem);
1603 gvn().set_type_bottom(mergemem);
1604 map()->set_memory(mergemem);
1605 }
1606
1607 //------------------------------set_all_memory_call----------------------------
1608 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1609 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1610 set_all_memory(newmem);
1611 }
1612
1613 //=============================================================================
1614 //
1615 // parser factory methods for MemNodes
1616 //
1617 // These are layered on top of the factory methods in LoadNode and StoreNode,
1618 // and integrate with the parser's memory state and _gvn engine.
1619 //
1620
1621 // factory methods in "int adr_idx"
1622 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1623 MemNode::MemOrd mo,
1624 LoadNode::ControlDependency control_dependency,
1625 bool require_atomic_access,
1626 bool unaligned,
1627 bool mismatched,
1628 bool unsafe,
1629 uint8_t barrier_data) {
1630 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1631 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1632 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1633 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1634 Node* mem = memory(adr_idx);
1635 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1636 ld = _gvn.transform(ld);
1637 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1638 // Improve graph before escape analysis and boxing elimination.
1639 record_for_igvn(ld);
1640 if (ld->is_DecodeN()) {
1641 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1642 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1643 // a Phi). Recording such cases is still perfectly sound, but may be
1644 // unnecessary and result in some minor IGVN overhead.
1645 record_for_igvn(ld->in(1));
1646 }
1647 }
1648 return ld;
1649 }
1650
1651 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1652 MemNode::MemOrd mo,
1653 bool require_atomic_access,
1654 bool unaligned,
1655 bool mismatched,
1656 bool unsafe,
1670 if (unsafe) {
1671 st->as_Store()->set_unsafe_access();
1672 }
1673 st->as_Store()->set_barrier_data(barrier_data);
1674 st = _gvn.transform(st);
1675 set_memory(st, adr_idx);
1676 // Back-to-back stores can only remove intermediate store with DU info
1677 // so push on worklist for optimizer.
1678 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1679 record_for_igvn(st);
1680
1681 return st;
1682 }
1683
1684 Node* GraphKit::access_store_at(Node* obj,
1685 Node* adr,
1686 const TypePtr* adr_type,
1687 Node* val,
1688 const Type* val_type,
1689 BasicType bt,
1690 DecoratorSet decorators) {
1691 // Transformation of a value which could be null pointer (CastPP #null)
1692 // could be delayed during Parse (for example, in adjust_map_after_if()).
1693 // Execute transformation here to avoid barrier generation in such case.
1694 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1695 val = _gvn.makecon(TypePtr::NULL_PTR);
1696 }
1697
1698 if (stopped()) {
1699 return top(); // Dead path ?
1700 }
1701
1702 assert(val != nullptr, "not dead path");
1703
1704 C2AccessValuePtr addr(adr, adr_type);
1705 C2AccessValue value(val, val_type);
1706 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr);
1707 if (access.is_raw()) {
1708 return _barrier_set->BarrierSetC2::store_at(access, value);
1709 } else {
1710 return _barrier_set->store_at(access, value);
1711 }
1712 }
1713
1714 Node* GraphKit::access_load_at(Node* obj, // containing obj
1715 Node* adr, // actual address to store val at
1716 const TypePtr* adr_type,
1717 const Type* val_type,
1718 BasicType bt,
1719 DecoratorSet decorators) {
1720 if (stopped()) {
1721 return top(); // Dead path ?
1722 }
1723
1724 SavedState old_state(this);
1725 C2AccessValuePtr addr(adr, adr_type);
1726 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr);
1727 Node* load;
1728 if (access.is_raw()) {
1729 load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1730 } else {
1731 load = _barrier_set->load_at(access, val_type);
1732 }
1733
1734 // Restore the previous state only if the load got folded to a constant
1735 // and we can discard any barriers that might have been added.
1736 if (load == nullptr || !load->is_Con()) {
1737 old_state.discard();
1738 }
1739 return load;
1740 }
1741
1742 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1743 const Type* val_type,
1744 BasicType bt,
1745 DecoratorSet decorators) {
1746 if (stopped()) {
1828 Node* new_val,
1829 const Type* value_type,
1830 BasicType bt,
1831 DecoratorSet decorators) {
1832 C2AccessValuePtr addr(adr, adr_type);
1833 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1834 if (access.is_raw()) {
1835 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1836 } else {
1837 return _barrier_set->atomic_add_at(access, new_val, value_type);
1838 }
1839 }
1840
1841 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1842 return _barrier_set->clone(this, src, dst, size, is_array);
1843 }
1844
1845 //-------------------------array_element_address-------------------------
1846 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1847 const TypeInt* sizetype, Node* ctrl) {
1848 uint shift = exact_log2(type2aelembytes(elembt));
1849 uint header = arrayOopDesc::base_offset_in_bytes(elembt);
1850
1851 // short-circuit a common case (saves lots of confusing waste motion)
1852 jint idx_con = find_int_con(idx, -1);
1853 if (idx_con >= 0) {
1854 intptr_t offset = header + ((intptr_t)idx_con << shift);
1855 return basic_plus_adr(ary, offset);
1856 }
1857
1858 // must be correct type for alignment purposes
1859 Node* base = basic_plus_adr(ary, header);
1860 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1861 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1862 return basic_plus_adr(ary, base, scale);
1863 }
1864
1865 //-------------------------load_array_element-------------------------
1866 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
1867 const Type* elemtype = arytype->elem();
1868 BasicType elembt = elemtype->array_element_basic_type();
1869 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
1870 if (elembt == T_NARROWOOP) {
1871 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
1872 }
1873 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
1874 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
1875 return ld;
1876 }
1877
1878 //-------------------------set_arguments_for_java_call-------------------------
1879 // Arguments (pre-popped from the stack) are taken from the JVMS.
1880 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
1881 // Add the call arguments:
1882 uint nargs = call->method()->arg_size();
1883 for (uint i = 0; i < nargs; i++) {
1884 Node* arg = argument(i);
1885 call->init_req(i + TypeFunc::Parms, arg);
1886 }
1887 }
1888
1889 //---------------------------set_edges_for_java_call---------------------------
1890 // Connect a newly created call into the current JVMS.
1891 // A return value node (if any) is returned from set_edges_for_java_call.
1892 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
1893
1894 // Add the predefined inputs:
1895 call->init_req( TypeFunc::Control, control() );
1896 call->init_req( TypeFunc::I_O , i_o() );
1897 call->init_req( TypeFunc::Memory , reset_memory() );
1898 call->init_req( TypeFunc::FramePtr, frameptr() );
1899 call->init_req( TypeFunc::ReturnAdr, top() );
1900
1901 add_safepoint_edges(call, must_throw);
1902
1903 Node* xcall = _gvn.transform(call);
1904
1905 if (xcall == top()) {
1906 set_control(top());
1907 return;
1908 }
1909 assert(xcall == call, "call identity is stable");
1910
1911 // Re-use the current map to produce the result.
1912
1913 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
1914 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
1915 set_all_memory_call(xcall, separate_io_proj);
1916
1917 //return xcall; // no need, caller already has it
1918 }
1919
1920 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
1921 if (stopped()) return top(); // maybe the call folded up?
1922
1923 // Capture the return value, if any.
1924 Node* ret;
1925 if (call->method() == nullptr ||
1926 call->method()->return_type()->basic_type() == T_VOID)
1927 ret = top();
1928 else ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
1929
1930 // Note: Since any out-of-line call can produce an exception,
1931 // we always insert an I_O projection from the call into the result.
1932
1933 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
1934
1935 if (separate_io_proj) {
1936 // The caller requested separate projections be used by the fall
1937 // through and exceptional paths, so replace the projections for
1938 // the fall through path.
1939 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
1940 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
1941 }
1942 return ret;
1943 }
1944
1945 //--------------------set_predefined_input_for_runtime_call--------------------
1946 // Reading and setting the memory state is way conservative here.
1947 // The real problem is that I am not doing real Type analysis on memory,
1948 // so I cannot distinguish card mark stores from other stores. Across a GC
1949 // point the Store Barrier and the card mark memory has to agree. I cannot
1950 // have a card mark store and its barrier split across the GC point from
1951 // either above or below. Here I get that to happen by reading ALL of memory.
1952 // A better answer would be to separate out card marks from other memory.
1953 // For now, return the input memory state, so that it can be reused
1954 // after the call, if this call has restricted memory effects.
1955 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
1956 // Set fixed predefined input arguments
1957 call->init_req(TypeFunc::Control, control());
1958 call->init_req(TypeFunc::I_O, top()); // does no i/o
1959 call->init_req(TypeFunc::ReturnAdr, top());
1960 if (call->is_CallLeafPure()) {
1961 call->init_req(TypeFunc::Memory, top());
2023 if (use->is_MergeMem()) {
2024 wl.push(use);
2025 }
2026 }
2027 }
2028
2029 // Replace the call with the current state of the kit.
2030 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2031 JVMState* ejvms = nullptr;
2032 if (has_exceptions()) {
2033 ejvms = transfer_exceptions_into_jvms();
2034 }
2035
2036 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2037 ReplacedNodes replaced_nodes_exception;
2038 Node* ex_ctl = top();
2039
2040 SafePointNode* final_state = stop();
2041
2042 // Find all the needed outputs of this call
2043 CallProjections callprojs;
2044 call->extract_projections(&callprojs, true, do_asserts);
2045
2046 Unique_Node_List wl;
2047 Node* init_mem = call->in(TypeFunc::Memory);
2048 Node* final_mem = final_state->in(TypeFunc::Memory);
2049 Node* final_ctl = final_state->in(TypeFunc::Control);
2050 Node* final_io = final_state->in(TypeFunc::I_O);
2051
2052 // Replace all the old call edges with the edges from the inlining result
2053 if (callprojs.fallthrough_catchproj != nullptr) {
2054 C->gvn_replace_by(callprojs.fallthrough_catchproj, final_ctl);
2055 }
2056 if (callprojs.fallthrough_memproj != nullptr) {
2057 if (final_mem->is_MergeMem()) {
2058 // Parser's exits MergeMem was not transformed but may be optimized
2059 final_mem = _gvn.transform(final_mem);
2060 }
2061 C->gvn_replace_by(callprojs.fallthrough_memproj, final_mem);
2062 add_mergemem_users_to_worklist(wl, final_mem);
2063 }
2064 if (callprojs.fallthrough_ioproj != nullptr) {
2065 C->gvn_replace_by(callprojs.fallthrough_ioproj, final_io);
2066 }
2067
2068 // Replace the result with the new result if it exists and is used
2069 if (callprojs.resproj != nullptr && result != nullptr) {
2070 C->gvn_replace_by(callprojs.resproj, result);
2071 }
2072
2073 if (ejvms == nullptr) {
2074 // No exception edges to simply kill off those paths
2075 if (callprojs.catchall_catchproj != nullptr) {
2076 C->gvn_replace_by(callprojs.catchall_catchproj, C->top());
2077 }
2078 if (callprojs.catchall_memproj != nullptr) {
2079 C->gvn_replace_by(callprojs.catchall_memproj, C->top());
2080 }
2081 if (callprojs.catchall_ioproj != nullptr) {
2082 C->gvn_replace_by(callprojs.catchall_ioproj, C->top());
2083 }
2084 // Replace the old exception object with top
2085 if (callprojs.exobj != nullptr) {
2086 C->gvn_replace_by(callprojs.exobj, C->top());
2087 }
2088 } else {
2089 GraphKit ekit(ejvms);
2090
2091 // Load my combined exception state into the kit, with all phis transformed:
2092 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2093 replaced_nodes_exception = ex_map->replaced_nodes();
2094
2095 Node* ex_oop = ekit.use_exception_state(ex_map);
2096
2097 if (callprojs.catchall_catchproj != nullptr) {
2098 C->gvn_replace_by(callprojs.catchall_catchproj, ekit.control());
2099 ex_ctl = ekit.control();
2100 }
2101 if (callprojs.catchall_memproj != nullptr) {
2102 Node* ex_mem = ekit.reset_memory();
2103 C->gvn_replace_by(callprojs.catchall_memproj, ex_mem);
2104 add_mergemem_users_to_worklist(wl, ex_mem);
2105 }
2106 if (callprojs.catchall_ioproj != nullptr) {
2107 C->gvn_replace_by(callprojs.catchall_ioproj, ekit.i_o());
2108 }
2109
2110 // Replace the old exception object with the newly created one
2111 if (callprojs.exobj != nullptr) {
2112 C->gvn_replace_by(callprojs.exobj, ex_oop);
2113 }
2114 }
2115
2116 // Disconnect the call from the graph
2117 call->disconnect_inputs(C);
2118 C->gvn_replace_by(call, C->top());
2119
2120 // Clean up any MergeMems that feed other MergeMems since the
2121 // optimizer doesn't like that.
2122 while (wl.size() > 0) {
2123 _gvn.transform(wl.pop());
2124 }
2125
2126 if (callprojs.fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2127 replaced_nodes.apply(C, final_ctl);
2128 }
2129 if (!ex_ctl->is_top() && do_replaced_nodes) {
2130 replaced_nodes_exception.apply(C, ex_ctl);
2131 }
2132 }
2133
2134
2135 //------------------------------increment_counter------------------------------
2136 // for statistics: increment a VM counter by 1
2137
2138 void GraphKit::increment_counter(address counter_addr) {
2139 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2140 increment_counter(adr1);
2141 }
2142
2143 void GraphKit::increment_counter(Node* counter_addr) {
2144 Node* ctrl = control();
2145 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2146 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2147 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2148 }
2149
2150 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2151 Node* halt = new HaltNode(ctrl, frameptr, reason
2152 PRODUCT_ONLY(COMMA generate_code_in_product));
2153 halt = _gvn.transform(halt);
2154 root()->add_req(halt);
2155 }
2156
2157 //------------------------------uncommon_trap----------------------------------
2158 // Bail out to the interpreter in mid-method. Implemented by calling the
2159 // uncommon_trap blob. This helper function inserts a runtime call with the
2160 // right debug info.
2161 Node* GraphKit::uncommon_trap(int trap_request,
2162 ciKlass* klass, const char* comment,
2163 bool must_throw,
2164 bool keep_exact_action) {
2165 if (failing_internal()) {
2166 stop();
2167 }
2168 if (stopped()) return nullptr; // trap reachable?
2169
2170 // Note: If ProfileTraps is true, and if a deopt. actually
2171 // occurs here, the runtime will make sure an MDO exists. There is
2172 // no need to call method()->ensure_method_data() at this point.
2173
2174 // Set the stack pointer to the right value for reexecution:
2316 *
2317 * @param n node that the type applies to
2318 * @param exact_kls type from profiling
2319 * @param maybe_null did profiling see null?
2320 *
2321 * @return node with improved type
2322 */
2323 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2324 const Type* current_type = _gvn.type(n);
2325 assert(UseTypeSpeculation, "type speculation must be on");
2326
2327 const TypePtr* speculative = current_type->speculative();
2328
2329 // Should the klass from the profile be recorded in the speculative type?
2330 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2331 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2332 const TypeOopPtr* xtype = tklass->as_instance_type();
2333 assert(xtype->klass_is_exact(), "Should be exact");
2334 // Any reason to believe n is not null (from this profiling or a previous one)?
2335 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2336 const TypePtr* ptr = (ptr_kind == ProfileMaybeNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2337 // record the new speculative type's depth
2338 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2339 speculative = speculative->with_inline_depth(jvms()->depth());
2340 } else if (current_type->would_improve_ptr(ptr_kind)) {
2341 // Profiling report that null was never seen so we can change the
2342 // speculative type to non null ptr.
2343 if (ptr_kind == ProfileAlwaysNull) {
2344 speculative = TypePtr::NULL_PTR;
2345 } else {
2346 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2347 const TypePtr* ptr = TypePtr::NOTNULL;
2348 if (speculative != nullptr) {
2349 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2350 } else {
2351 speculative = ptr;
2352 }
2353 }
2354 }
2355
2356 if (speculative != current_type->speculative()) {
2357 // Build a type with a speculative type (what we think we know
2358 // about the type but will need a guard when we use it)
2359 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::OffsetBot, TypeOopPtr::InstanceBot, speculative);
2360 // We're changing the type, we need a new CheckCast node to carry
2361 // the new type. The new type depends on the control: what
2362 // profiling tells us is only valid from here as far as we can
2363 // tell.
2364 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2365 cast = _gvn.transform(cast);
2366 replace_in_map(n, cast);
2367 n = cast;
2368 }
2369
2370 return n;
2371 }
2372
2373 /**
2374 * Record profiling data from receiver profiling at an invoke with the
2375 * type system so that it can propagate it (speculation)
2376 *
2377 * @param n receiver node
2378 *
2379 * @return node with improved type
2380 */
2381 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2382 if (!UseTypeSpeculation) {
2383 return n;
2384 }
2385 ciKlass* exact_kls = profile_has_unique_klass();
2386 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2387 if ((java_bc() == Bytecodes::_checkcast ||
2388 java_bc() == Bytecodes::_instanceof ||
2389 java_bc() == Bytecodes::_aastore) &&
2390 method()->method_data()->is_mature()) {
2391 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2392 if (data != nullptr) {
2393 if (!data->as_BitData()->null_seen()) {
2394 ptr_kind = ProfileNeverNull;
2395 } else {
2396 if (TypeProfileCasts) {
2397 assert(data->is_ReceiverTypeData(), "bad profile data type");
2398 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2399 uint i = 0;
2400 for (; i < call->row_limit(); i++) {
2401 ciKlass* receiver = call->receiver(i);
2402 if (receiver != nullptr) {
2403 break;
2404 }
2405 }
2406 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2407 }
2408 }
2409 }
2410 }
2411 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2412 }
2413
2414 /**
2415 * Record profiling data from argument profiling at an invoke with the
2416 * type system so that it can propagate it (speculation)
2417 *
2418 * @param dest_method target method for the call
2419 * @param bc what invoke bytecode is this?
2420 */
2421 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2422 if (!UseTypeSpeculation) {
2423 return;
2424 }
2425 const TypeFunc* tf = TypeFunc::make(dest_method);
2426 int nargs = tf->domain()->cnt() - TypeFunc::Parms;
2427 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2428 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2429 const Type *targ = tf->domain()->field_at(j + TypeFunc::Parms);
2430 if (is_reference_type(targ->basic_type())) {
2431 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2432 ciKlass* better_type = nullptr;
2433 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2434 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2435 }
2436 i++;
2437 }
2438 }
2439 }
2440
2441 /**
2442 * Record profiling data from parameter profiling at an invoke with
2443 * the type system so that it can propagate it (speculation)
2444 */
2445 void GraphKit::record_profiled_parameters_for_speculation() {
2446 if (!UseTypeSpeculation) {
2447 return;
2448 }
2449 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2569 // The first null ends the list.
2570 Node* parm0, Node* parm1,
2571 Node* parm2, Node* parm3,
2572 Node* parm4, Node* parm5,
2573 Node* parm6, Node* parm7) {
2574 assert(call_addr != nullptr, "must not call null targets");
2575
2576 // Slow-path call
2577 bool is_leaf = !(flags & RC_NO_LEAF);
2578 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2579 if (call_name == nullptr) {
2580 assert(!is_leaf, "must supply name for leaf");
2581 call_name = OptoRuntime::stub_name(call_addr);
2582 }
2583 CallNode* call;
2584 if (!is_leaf) {
2585 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2586 } else if (flags & RC_NO_FP) {
2587 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2588 } else if (flags & RC_VECTOR){
2589 uint num_bits = call_type->range()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2590 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2591 } else if (flags & RC_PURE) {
2592 assert(adr_type == nullptr, "pure call does not touch memory");
2593 call = new CallLeafPureNode(call_type, call_addr, call_name);
2594 } else {
2595 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2596 }
2597
2598 // The following is similar to set_edges_for_java_call,
2599 // except that the memory effects of the call are restricted to AliasIdxRaw.
2600
2601 // Slow path call has no side-effects, uses few values
2602 bool wide_in = !(flags & RC_NARROW_MEM);
2603 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2604
2605 Node* prev_mem = nullptr;
2606 if (wide_in) {
2607 prev_mem = set_predefined_input_for_runtime_call(call);
2608 } else {
2609 assert(!wide_out, "narrow in => narrow out");
2610 Node* narrow_mem = memory(adr_type);
2611 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2612 }
2613
2614 // Hook each parm in order. Stop looking at the first null.
2615 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2616 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2617 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2618 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2619 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2620 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2621 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2622 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2623 /* close each nested if ===> */ } } } } } } } }
2624 assert(call->in(call->req()-1) != nullptr, "must initialize all parms");
2625
2626 if (!is_leaf) {
2627 // Non-leaves can block and take safepoints:
2628 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2629 }
2630 // Non-leaves can throw exceptions:
2631 if (has_io) {
2632 call->set_req(TypeFunc::I_O, i_o());
2633 }
2634
2635 if (flags & RC_UNCOMMON) {
2636 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2637 // (An "if" probability corresponds roughly to an unconditional count.
2638 // Sort of.)
2639 call->set_cnt(PROB_UNLIKELY_MAG(4));
2640 }
2641
2642 Node* c = _gvn.transform(call);
2643 assert(c == call, "cannot disappear");
2644
2652
2653 if (has_io) {
2654 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2655 }
2656 return call;
2657
2658 }
2659
2660 // i2b
2661 Node* GraphKit::sign_extend_byte(Node* in) {
2662 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2663 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2664 }
2665
2666 // i2s
2667 Node* GraphKit::sign_extend_short(Node* in) {
2668 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2669 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2670 }
2671
2672 //------------------------------merge_memory-----------------------------------
2673 // Merge memory from one path into the current memory state.
2674 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2675 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2676 Node* old_slice = mms.force_memory();
2677 Node* new_slice = mms.memory2();
2678 if (old_slice != new_slice) {
2679 PhiNode* phi;
2680 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2681 if (mms.is_empty()) {
2682 // clone base memory Phi's inputs for this memory slice
2683 assert(old_slice == mms.base_memory(), "sanity");
2684 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2685 _gvn.set_type(phi, Type::MEMORY);
2686 for (uint i = 1; i < phi->req(); i++) {
2687 phi->init_req(i, old_slice->in(i));
2688 }
2689 } else {
2690 phi = old_slice->as_Phi(); // Phi was generated already
2691 }
2748 gvn.transform(iff);
2749 if (!bol->is_Con()) gvn.record_for_igvn(iff);
2750 return iff;
2751 }
2752
2753 //-------------------------------gen_subtype_check-----------------------------
2754 // Generate a subtyping check. Takes as input the subtype and supertype.
2755 // Returns 2 values: sets the default control() to the true path and returns
2756 // the false path. Only reads invariant memory; sets no (visible) memory.
2757 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
2758 // but that's not exposed to the optimizer. This call also doesn't take in an
2759 // Object; if you wish to check an Object you need to load the Object's class
2760 // prior to coming here.
2761 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
2762 ciMethod* method, int bci) {
2763 Compile* C = gvn.C;
2764 if ((*ctrl)->is_top()) {
2765 return C->top();
2766 }
2767
2768 // Fast check for identical types, perhaps identical constants.
2769 // The types can even be identical non-constants, in cases
2770 // involving Array.newInstance, Object.clone, etc.
2771 if (subklass == superklass)
2772 return C->top(); // false path is dead; no test needed.
2773
2774 if (gvn.type(superklass)->singleton()) {
2775 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2776 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
2777
2778 // In the common case of an exact superklass, try to fold up the
2779 // test before generating code. You may ask, why not just generate
2780 // the code and then let it fold up? The answer is that the generated
2781 // code will necessarily include null checks, which do not always
2782 // completely fold away. If they are also needless, then they turn
2783 // into a performance loss. Example:
2784 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
2785 // Here, the type of 'fa' is often exact, so the store check
2786 // of fa[1]=x will fold up, without testing the nullness of x.
2787 //
2788 // At macro expansion, we would have already folded the SubTypeCheckNode
2789 // being expanded here because we always perform the static sub type
2790 // check in SubTypeCheckNode::sub() regardless of whether
2791 // StressReflectiveCode is set or not. We can therefore skip this
2792 // static check when StressReflectiveCode is on.
2793 switch (C->static_subtype_check(superk, subk)) {
2794 case Compile::SSC_always_false:
2795 {
2796 Node* always_fail = *ctrl;
2797 *ctrl = gvn.C->top();
2798 return always_fail;
2799 }
2800 case Compile::SSC_always_true:
2801 return C->top();
2802 case Compile::SSC_easy_test:
2803 {
2804 // Just do a direct pointer compare and be done.
2805 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
2806 *ctrl = gvn.transform(new IfTrueNode(iff));
2807 return gvn.transform(new IfFalseNode(iff));
2808 }
2809 case Compile::SSC_full_test:
2810 break;
2811 default:
2812 ShouldNotReachHere();
2813 }
2814 }
2815
2816 // %%% Possible further optimization: Even if the superklass is not exact,
2817 // if the subklass is the unique subtype of the superklass, the check
2818 // will always succeed. We could leave a dependency behind to ensure this.
2819
2820 // First load the super-klass's check-offset
2821 Node* p1 = gvn.transform(AddPNode::make_off_heap(superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
2822 Node* m = C->immutable_memory();
2823 Node* chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
2824 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
2825 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
2863 gvn.record_for_igvn(r_ok_subtype);
2864
2865 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
2866 // SubTypeCheck node
2867 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
2868 ciCallProfile profile = method->call_profile_at_bci(bci);
2869 float total_prob = 0;
2870 for (int i = 0; profile.has_receiver(i); ++i) {
2871 float prob = profile.receiver_prob(i);
2872 total_prob += prob;
2873 }
2874 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
2875 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
2876 for (int i = 0; profile.has_receiver(i); ++i) {
2877 ciKlass* klass = profile.receiver(i);
2878 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
2879 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
2880 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
2881 continue;
2882 }
2883 float prob = profile.receiver_prob(i);
2884 ConNode* klass_node = gvn.makecon(klass_t);
2885 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
2886 Node* iftrue = gvn.transform(new IfTrueNode(iff));
2887
2888 if (result == Compile::SSC_always_true) {
2889 r_ok_subtype->add_req(iftrue);
2890 } else {
2891 assert(result == Compile::SSC_always_false, "");
2892 r_not_subtype->add_req(iftrue);
2893 }
2894 *ctrl = gvn.transform(new IfFalseNode(iff));
2895 }
2896 }
2897 }
2898
2899 // See if we get an immediate positive hit. Happens roughly 83% of the
2900 // time. Test to see if the value loaded just previously from the subklass
2901 // is exactly the superklass.
2902 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
2916 igvn->remove_globally_dead_node(r_not_subtype, PhaseIterGVN::NodeOrigin::Speculative);
2917 }
2918 return not_subtype_ctrl;
2919 }
2920
2921 r_ok_subtype->init_req(1, iftrue1);
2922
2923 // Check for immediate negative hit. Happens roughly 11% of the time (which
2924 // is roughly 63% of the remaining cases). Test to see if the loaded
2925 // check-offset points into the subklass display list or the 1-element
2926 // cache. If it points to the display (and NOT the cache) and the display
2927 // missed then it's not a subtype.
2928 Node *cacheoff = gvn.intcon(cacheoff_con);
2929 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
2930 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
2931 *ctrl = gvn.transform(new IfFalseNode(iff2));
2932
2933 // Check for self. Very rare to get here, but it is taken 1/3 the time.
2934 // No performance impact (too rare) but allows sharing of secondary arrays
2935 // which has some footprint reduction.
2936 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
2937 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
2938 *ctrl = gvn.transform(new IfFalseNode(iff3));
2939
2940 // -- Roads not taken here: --
2941 // We could also have chosen to perform the self-check at the beginning
2942 // of this code sequence, as the assembler does. This would not pay off
2943 // the same way, since the optimizer, unlike the assembler, can perform
2944 // static type analysis to fold away many successful self-checks.
2945 // Non-foldable self checks work better here in second position, because
2946 // the initial primary superclass check subsumes a self-check for most
2947 // types. An exception would be a secondary type like array-of-interface,
2948 // which does not appear in its own primary supertype display.
2949 // Finally, we could have chosen to move the self-check into the
2950 // PartialSubtypeCheckNode, and from there out-of-line in a platform
2951 // dependent manner. But it is worthwhile to have the check here,
2952 // where it can be perhaps be optimized. The cost in code space is
2953 // small (register compare, branch).
2954
2955 // Now do a linear scan of the secondary super-klass array. Again, no real
2956 // performance impact (too rare) but it's gotta be done.
2957 // Since the code is rarely used, there is no penalty for moving it
2958 // out of line, and it can only improve I-cache density.
2959 // The decision to inline or out-of-line this final check is platform
2960 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
2961 Node* psc = gvn.transform(
2962 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
2963
2964 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
2965 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
2966 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
2967
2968 // Return false path; set default control to true path.
2969 *ctrl = gvn.transform(r_ok_subtype);
2970 return gvn.transform(r_not_subtype);
2971 }
2972
2973 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
2974 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
2975 if (expand_subtype_check) {
2976 MergeMemNode* mem = merged_memory();
2977 Node* ctrl = control();
2978 Node* subklass = obj_or_subklass;
2979 if (!_gvn.type(obj_or_subklass)->isa_klassptr()) {
2980 subklass = load_object_klass(obj_or_subklass);
2981 }
2982
2983 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
2984 set_control(ctrl);
2985 return n;
2986 }
2987
2988 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
2989 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
2990 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
2991 set_control(_gvn.transform(new IfTrueNode(iff)));
2992 return _gvn.transform(new IfFalseNode(iff));
2993 }
2994
2995 // Profile-driven exact type check:
2996 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
2997 float prob,
2998 Node* *casted_receiver) {
2999 assert(!klass->is_interface(), "no exact type check on interfaces");
3000
3001 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3002 Node* recv_klass = load_object_klass(receiver);
3003 Node* want_klass = makecon(tklass);
3004 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3005 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3006 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3007 set_control( _gvn.transform(new IfTrueNode (iff)));
3008 Node* fail = _gvn.transform(new IfFalseNode(iff));
3009
3010 if (!stopped()) {
3011 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3012 const TypeOopPtr* recvx_type = tklass->as_instance_type();
3013 assert(recvx_type->klass_is_exact(), "");
3014
3015 if (!receiver_type->higher_equal(recvx_type)) { // ignore redundant casts
3016 // Subsume downstream occurrences of receiver with a cast to
3017 // recv_xtype, since now we know what the type will be.
3018 Node* cast = new CheckCastPPNode(control(), receiver, recvx_type);
3019 (*casted_receiver) = _gvn.transform(cast);
3020 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3021 // (User must make the replace_in_map call.)
3022 }
3023 }
3024
3025 return fail;
3026 }
3027
3028 //------------------------------subtype_check_receiver-------------------------
3029 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3030 Node** casted_receiver) {
3031 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3032 Node* want_klass = makecon(tklass);
3033
3034 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3035
3036 // Ignore interface type information until interface types are properly tracked.
3037 if (!stopped() && !klass->is_interface()) {
3038 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3039 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3040 if (!receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3041 Node* cast = new CheckCastPPNode(control(), receiver, recv_type);
3042 (*casted_receiver) = _gvn.transform(cast);
3043 }
3044 }
3045
3046 return slow_ctl;
3047 }
3048
3049 //------------------------------seems_never_null-------------------------------
3050 // Use null_seen information if it is available from the profile.
3051 // If we see an unexpected null at a type check we record it and force a
3052 // recompile; the offending check will be recompiled to handle nulls.
3053 // If we see several offending BCIs, then all checks in the
3054 // method will be recompiled.
3055 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3056 speculating = !_gvn.type(obj)->speculative_maybe_null();
3057 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3058 if (UncommonNullCast // Cutout for this technique
3059 && obj != null() // And not the -Xcomp stupid case?
3060 && !too_many_traps(reason)
3061 ) {
3062 if (speculating) {
3131
3132 //------------------------maybe_cast_profiled_receiver-------------------------
3133 // If the profile has seen exactly one type, narrow to exactly that type.
3134 // Subsequent type checks will always fold up.
3135 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3136 const TypeKlassPtr* require_klass,
3137 ciKlass* spec_klass,
3138 bool safe_for_replace) {
3139 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3140
3141 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3142
3143 // Make sure we haven't already deoptimized from this tactic.
3144 if (too_many_traps_or_recompiles(reason))
3145 return nullptr;
3146
3147 // (No, this isn't a call, but it's enough like a virtual call
3148 // to use the same ciMethod accessor to get the profile info...)
3149 // If we have a speculative type use it instead of profiling (which
3150 // may not help us)
3151 ciKlass* exact_kls = spec_klass == nullptr ? profile_has_unique_klass() : spec_klass;
3152 if (exact_kls != nullptr) {// no cast failures here
3153 if (require_klass == nullptr ||
3154 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3155 // If we narrow the type to match what the type profile sees or
3156 // the speculative type, we can then remove the rest of the
3157 // cast.
3158 // This is a win, even if the exact_kls is very specific,
3159 // because downstream operations, such as method calls,
3160 // will often benefit from the sharper type.
3161 Node* exact_obj = not_null_obj; // will get updated in place...
3162 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3163 &exact_obj);
3164 { PreserveJVMState pjvms(this);
3165 set_control(slow_ctl);
3166 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3167 }
3168 if (safe_for_replace) {
3169 replace_in_map(not_null_obj, exact_obj);
3170 }
3171 return exact_obj;
3261 // If not_null_obj is dead, only null-path is taken
3262 if (stopped()) { // Doing instance-of on a null?
3263 set_control(null_ctl);
3264 return intcon(0);
3265 }
3266 region->init_req(_null_path, null_ctl);
3267 phi ->init_req(_null_path, intcon(0)); // Set null path value
3268 if (null_ctl == top()) {
3269 // Do this eagerly, so that pattern matches like is_diamond_phi
3270 // will work even during parsing.
3271 assert(_null_path == PATH_LIMIT-1, "delete last");
3272 region->del_req(_null_path);
3273 phi ->del_req(_null_path);
3274 }
3275
3276 // Do we know the type check always succeed?
3277 bool known_statically = false;
3278 if (_gvn.type(superklass)->singleton()) {
3279 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3280 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3281 if (subk->is_loaded()) {
3282 int static_res = C->static_subtype_check(superk, subk);
3283 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3284 }
3285 }
3286
3287 if (!known_statically) {
3288 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3289 // We may not have profiling here or it may not help us. If we
3290 // have a speculative type use it to perform an exact cast.
3291 ciKlass* spec_obj_type = obj_type->speculative_type();
3292 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3293 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3294 if (stopped()) { // Profile disagrees with this path.
3295 set_control(null_ctl); // Null is the only remaining possibility.
3296 return intcon(0);
3297 }
3298 if (cast_obj != nullptr) {
3299 not_null_obj = cast_obj;
3300 }
3301 }
3317 record_for_igvn(region);
3318
3319 // If we know the type check always succeeds then we don't use the
3320 // profiling data at this bytecode. Don't lose it, feed it to the
3321 // type system as a speculative type.
3322 if (safe_for_replace) {
3323 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3324 replace_in_map(obj, casted_obj);
3325 }
3326
3327 return _gvn.transform(phi);
3328 }
3329
3330 //-------------------------------gen_checkcast---------------------------------
3331 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3332 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3333 // uncommon-trap paths work. Adjust stack after this call.
3334 // If failure_control is supplied and not null, it is filled in with
3335 // the control edge for the cast failure. Otherwise, an appropriate
3336 // uncommon trap or exception is thrown.
3337 Node* GraphKit::gen_checkcast(Node *obj, Node* superklass,
3338 Node* *failure_control) {
3339 kill_dead_locals(); // Benefit all the uncommon traps
3340 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3341 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3342 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3343
3344 // Fast cutout: Check the case that the cast is vacuously true.
3345 // This detects the common cases where the test will short-circuit
3346 // away completely. We do this before we perform the null check,
3347 // because if the test is going to turn into zero code, we don't
3348 // want a residual null check left around. (Causes a slowdown,
3349 // for example, in some objArray manipulations, such as a[i]=a[j].)
3350 if (improved_klass_ptr_type->singleton()) {
3351 const TypeOopPtr* objtp = _gvn.type(obj)->isa_oopptr();
3352 if (objtp != nullptr) {
3353 switch (C->static_subtype_check(improved_klass_ptr_type, objtp->as_klass_type())) {
3354 case Compile::SSC_always_true:
3355 // If we know the type check always succeed then we don't use
3356 // the profiling data at this bytecode. Don't lose it, feed it
3357 // to the type system as a speculative type.
3358 return record_profiled_receiver_for_speculation(obj);
3359 case Compile::SSC_always_false:
3360 // It needs a null check because a null will *pass* the cast check.
3361 // A non-null value will always produce an exception.
3362 if (!objtp->maybe_null()) {
3363 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3364 Deoptimization::DeoptReason reason = is_aastore ?
3365 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3366 builtin_throw(reason);
3367 return top();
3368 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3369 return null_assert(obj);
3370 }
3371 break; // Fall through to full check
3372 default:
3373 break;
3374 }
3375 }
3376 }
3377
3378 ciProfileData* data = nullptr;
3379 bool safe_for_replace = false;
3380 if (failure_control == nullptr) { // use MDO in regular case only
3381 assert(java_bc() == Bytecodes::_aastore ||
3382 java_bc() == Bytecodes::_checkcast,
3383 "interpreter profiles type checks only for these BCs");
3384 data = method()->method_data()->bci_to_data(bci());
3385 safe_for_replace = true;
3386 }
3387
3388 // Make the merge point
3389 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3390 RegionNode* region = new RegionNode(PATH_LIMIT);
3391 Node* phi = new PhiNode(region, toop);
3392 C->set_has_split_ifs(true); // Has chance for split-if optimization
3393
3394 // Use null-cast information if it is available
3395 bool speculative_not_null = false;
3396 bool never_see_null = ((failure_control == nullptr) // regular case only
3397 && seems_never_null(obj, data, speculative_not_null));
3398
3399 // Null check; get casted pointer; set region slot 3
3400 Node* null_ctl = top();
3401 Node* not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3402
3403 // If not_null_obj is dead, only null-path is taken
3404 if (stopped()) { // Doing instance-of on a null?
3405 set_control(null_ctl);
3406 return null();
3407 }
3408 region->init_req(_null_path, null_ctl);
3409 phi ->init_req(_null_path, null()); // Set null path value
3410 if (null_ctl == top()) {
3411 // Do this eagerly, so that pattern matches like is_diamond_phi
3412 // will work even during parsing.
3413 assert(_null_path == PATH_LIMIT-1, "delete last");
3414 region->del_req(_null_path);
3415 phi ->del_req(_null_path);
3416 }
3417
3418 Node* cast_obj = nullptr;
3419 if (improved_klass_ptr_type->klass_is_exact()) {
3420 // The following optimization tries to statically cast the speculative type of the object
3421 // (for example obtained during profiling) to the type of the superklass and then do a
3422 // dynamic check that the type of the object is what we expect. To work correctly
3423 // for checkcast and aastore the type of superklass should be exact.
3424 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3425 // We may not have profiling here or it may not help us. If we have
3426 // a speculative type use it to perform an exact cast.
3427 ciKlass* spec_obj_type = obj_type->speculative_type();
3428 if (spec_obj_type != nullptr || data != nullptr) {
3429 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3430 if (cast_obj != nullptr) {
3431 if (failure_control != nullptr) // failure is now impossible
3432 (*failure_control) = top();
3433 // adjust the type of the phi to the exact klass:
3434 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3435 }
3436 }
3437 }
3438
3439 if (cast_obj == nullptr) {
3440 // Generate the subtype check
3441 Node* improved_superklass = superklass;
3442 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3443 improved_superklass = makecon(improved_klass_ptr_type);
3444 }
3445 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3446
3447 // Plug in success path into the merge
3448 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3449 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3450 if (failure_control == nullptr) {
3451 if (not_subtype_ctrl != top()) { // If failure is possible
3452 PreserveJVMState pjvms(this);
3453 set_control(not_subtype_ctrl);
3454 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3455 Deoptimization::DeoptReason reason = is_aastore ?
3456 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3457 builtin_throw(reason);
3458 }
3459 } else {
3460 (*failure_control) = not_subtype_ctrl;
3461 }
3462 }
3463
3464 region->init_req(_obj_path, control());
3465 phi ->init_req(_obj_path, cast_obj);
3466
3467 // A merge of null or Casted-NotNull obj
3468 Node* res = _gvn.transform(phi);
3469
3470 // Note I do NOT always 'replace_in_map(obj,result)' here.
3471 // if( tk->klass()->can_be_primary_super() )
3472 // This means that if I successfully store an Object into an array-of-String
3473 // I 'forget' that the Object is really now known to be a String. I have to
3474 // do this because we don't have true union types for interfaces - if I store
3475 // a Baz into an array-of-Interface and then tell the optimizer it's an
3476 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3477 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3478 // replace_in_map( obj, res );
3479
3480 // Return final merged results
3481 set_control( _gvn.transform(region) );
3482 record_for_igvn(region);
3483
3484 return record_profiled_receiver_for_speculation(res);
3485 }
3486
3487 //------------------------------next_monitor-----------------------------------
3488 // What number should be given to the next monitor?
3489 int GraphKit::next_monitor() {
3490 int current = jvms()->monitor_depth()* C->sync_stack_slots();
3491 int next = current + C->sync_stack_slots();
3492 // Keep the toplevel high water mark current:
3493 if (C->fixed_slots() < next) C->set_fixed_slots(next);
3494 return current;
3495 }
3496
3497 //------------------------------insert_mem_bar---------------------------------
3498 // Memory barrier to avoid floating things around
3499 // The membar serves as a pinch point between both control and all memory slices.
3500 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
3501 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
3502 mb->init_req(TypeFunc::Control, control());
3503 mb->init_req(TypeFunc::Memory, reset_memory());
3504 Node* membar = _gvn.transform(mb);
3607 lock->create_lock_counter(map()->jvms());
3608 increment_counter(lock->counter()->addr());
3609 }
3610 #endif
3611
3612 return flock;
3613 }
3614
3615
3616 //------------------------------shared_unlock----------------------------------
3617 // Emit unlocking code.
3618 void GraphKit::shared_unlock(Node* box, Node* obj) {
3619 // bci is either a monitorenter bc or InvocationEntryBci
3620 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
3621 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
3622
3623 if (stopped()) { // Dead monitor?
3624 map()->pop_monitor(); // Kill monitor from debug info
3625 return;
3626 }
3627
3628 // Memory barrier to avoid floating things down past the locked region
3629 insert_mem_bar(Op_MemBarReleaseLock);
3630
3631 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
3632 UnlockNode *unlock = new UnlockNode(C, tf);
3633 #ifdef ASSERT
3634 unlock->set_dbg_jvms(sync_jvms());
3635 #endif
3636 uint raw_idx = Compile::AliasIdxRaw;
3637 unlock->init_req( TypeFunc::Control, control() );
3638 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
3639 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
3640 unlock->init_req( TypeFunc::FramePtr, frameptr() );
3641 unlock->init_req( TypeFunc::ReturnAdr, top() );
3642
3643 unlock->init_req(TypeFunc::Parms + 0, obj);
3644 unlock->init_req(TypeFunc::Parms + 1, box);
3645 unlock = _gvn.transform(unlock)->as_Unlock();
3646
3647 Node* mem = reset_memory();
3648
3649 // unlock has no side-effects, sets few values
3650 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
3651
3652 // Kill monitor from debug info
3653 map()->pop_monitor( );
3654 }
3655
3656 //-------------------------------get_layout_helper-----------------------------
3657 // If the given klass is a constant or known to be an array,
3658 // fetch the constant layout helper value into constant_value
3659 // and return null. Otherwise, load the non-constant
3660 // layout helper value, and return the node which represents it.
3661 // This two-faced routine is useful because allocation sites
3662 // almost always feature constant types.
3663 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
3664 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
3665 if (!StressReflectiveCode && klass_t != nullptr) {
3666 bool xklass = klass_t->klass_is_exact();
3667 if (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM)) {
3668 jint lhelper;
3669 if (klass_t->isa_aryklassptr()) {
3670 BasicType elem = klass_t->as_instance_type()->isa_aryptr()->elem()->array_element_basic_type();
3671 if (is_reference_type(elem, true)) {
3672 elem = T_OBJECT;
3673 }
3674 lhelper = Klass::array_layout_helper(elem);
3675 } else {
3676 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
3677 }
3678 if (lhelper != Klass::_lh_neutral_value) {
3679 constant_value = lhelper;
3680 return (Node*) nullptr;
3681 }
3682 }
3683 }
3684 constant_value = Klass::_lh_neutral_value; // put in a known value
3685 Node* lhp = off_heap_plus_addr(klass_node, in_bytes(Klass::layout_helper_offset()));
3686 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
3687 }
3688
3689 // We just put in an allocate/initialize with a big raw-memory effect.
3690 // Hook selected additional alias categories on the initialization.
3691 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
3692 MergeMemNode* init_in_merge,
3693 Node* init_out_raw) {
3694 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
3695 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
3696
3697 Node* prevmem = kit.memory(alias_idx);
3698 init_in_merge->set_memory_at(alias_idx, prevmem);
3699 kit.set_memory(init_out_raw, alias_idx);
3700 }
3701
3702 //---------------------------set_output_for_allocation-------------------------
3703 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
3704 const TypeOopPtr* oop_type,
3705 bool deoptimize_on_exception) {
3706 int rawidx = Compile::AliasIdxRaw;
3707 alloc->set_req( TypeFunc::FramePtr, frameptr() );
3708 add_safepoint_edges(alloc);
3709 Node* allocx = _gvn.transform(alloc);
3710 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
3711 // create memory projection for i_o
3712 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
3713 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
3714
3715 // create a memory projection as for the normal control path
3716 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
3717 set_memory(malloc, rawidx);
3718
3719 // a normal slow-call doesn't change i_o, but an allocation does
3720 // we create a separate i_o projection for the normal control path
3721 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
3722 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
3723
3724 // put in an initialization barrier
3725 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
3726 rawoop)->as_Initialize();
3727 assert(alloc->initialization() == init, "2-way macro link must work");
3728 assert(init ->allocation() == alloc, "2-way macro link must work");
3729 {
3730 // Extract memory strands which may participate in the new object's
3731 // initialization, and source them from the new InitializeNode.
3732 // This will allow us to observe initializations when they occur,
3733 // and link them properly (as a group) to the InitializeNode.
3734 assert(init->in(InitializeNode::Memory) == malloc, "");
3735 MergeMemNode* minit_in = MergeMemNode::make(malloc);
3736 init->set_req(InitializeNode::Memory, minit_in);
3737 record_for_igvn(minit_in); // fold it up later, if possible
3738 Node* minit_out = memory(rawidx);
3739 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
3740 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
3741 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
3742 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
3743 // multiple projections as a result.
3744 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
3745 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
3746 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
3747 if (oop_type->isa_aryptr()) {
3748 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
3749 int elemidx = C->get_alias_index(telemref);
3750 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(elemidx))));
3751 } else if (oop_type->isa_instptr()) {
3752 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
3753 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
3754 ciField* field = ik->nonstatic_field_at(i);
3755 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
3756 continue; // do not bother to track really large numbers of fields
3757 // Find (or create) the alias category for this field:
3758 int fieldidx = C->alias_type(field)->index();
3759 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
3760 }
3761 }
3762 }
3763
3764 // Cast raw oop to the real thing...
3765 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
3766 javaoop = _gvn.transform(javaoop);
3767 C->set_recent_alloc(control(), javaoop);
3768 assert(just_allocated_object(control()) == javaoop, "just allocated");
3769
3770 #ifdef ASSERT
3782 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
3783 }
3784 }
3785 #endif //ASSERT
3786
3787 return javaoop;
3788 }
3789
3790 //---------------------------new_instance--------------------------------------
3791 // This routine takes a klass_node which may be constant (for a static type)
3792 // or may be non-constant (for reflective code). It will work equally well
3793 // for either, and the graph will fold nicely if the optimizer later reduces
3794 // the type to a constant.
3795 // The optional arguments are for specialized use by intrinsics:
3796 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
3797 // - If 'return_size_val', report the total object size to the caller.
3798 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3799 Node* GraphKit::new_instance(Node* klass_node,
3800 Node* extra_slow_test,
3801 Node* *return_size_val,
3802 bool deoptimize_on_exception) {
3803 // Compute size in doublewords
3804 // The size is always an integral number of doublewords, represented
3805 // as a positive bytewise size stored in the klass's layout_helper.
3806 // The layout_helper also encodes (in a low bit) the need for a slow path.
3807 jint layout_con = Klass::_lh_neutral_value;
3808 Node* layout_val = get_layout_helper(klass_node, layout_con);
3809 int layout_is_con = (layout_val == nullptr);
3810
3811 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
3812 // Generate the initial go-slow test. It's either ALWAYS (return a
3813 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
3814 // case) a computed value derived from the layout_helper.
3815 Node* initial_slow_test = nullptr;
3816 if (layout_is_con) {
3817 assert(!StressReflectiveCode, "stress mode does not use these paths");
3818 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
3819 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
3820 } else { // reflective case
3821 // This reflective path is used by Unsafe.allocateInstance.
3822 // (It may be stress-tested by specifying StressReflectiveCode.)
3823 // Basically, we want to get into the VM is there's an illegal argument.
3824 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
3825 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
3826 if (extra_slow_test != intcon(0)) {
3827 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
3828 }
3829 // (Macro-expander will further convert this to a Bool, if necessary.)
3840
3841 // Clear the low bits to extract layout_helper_size_in_bytes:
3842 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
3843 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
3844 size = _gvn.transform( new AndXNode(size, mask) );
3845 }
3846 if (return_size_val != nullptr) {
3847 (*return_size_val) = size;
3848 }
3849
3850 // This is a precise notnull oop of the klass.
3851 // (Actually, it need not be precise if this is a reflective allocation.)
3852 // It's what we cast the result to.
3853 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
3854 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
3855 const TypeOopPtr* oop_type = tklass->as_instance_type();
3856
3857 // Now generate allocation code
3858
3859 // The entire memory state is needed for slow path of the allocation
3860 // since GC and deoptimization can happened.
3861 Node *mem = reset_memory();
3862 set_all_memory(mem); // Create new memory state
3863
3864 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
3865 control(), mem, i_o(),
3866 size, klass_node,
3867 initial_slow_test);
3868
3869 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
3870 }
3871
3872 //-------------------------------new_array-------------------------------------
3873 // helper for both newarray and anewarray
3874 // The 'length' parameter is (obviously) the length of the array.
3875 // The optional arguments are for specialized use by intrinsics:
3876 // - If 'return_size_val', report the non-padded array size (sum of header size
3877 // and array body) to the caller.
3878 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
3879 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
3880 Node* length, // number of array elements
3881 int nargs, // number of arguments to push back for uncommon trap
3882 Node* *return_size_val,
3883 bool deoptimize_on_exception) {
3884 jint layout_con = Klass::_lh_neutral_value;
3885 Node* layout_val = get_layout_helper(klass_node, layout_con);
3886 int layout_is_con = (layout_val == nullptr);
3887
3888 if (!layout_is_con && !StressReflectiveCode &&
3889 !too_many_traps(Deoptimization::Reason_class_check)) {
3890 // This is a reflective array creation site.
3891 // Optimistically assume that it is a subtype of Object[],
3892 // so that we can fold up all the address arithmetic.
3893 layout_con = Klass::array_layout_helper(T_OBJECT);
3894 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
3895 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
3896 { BuildCutout unless(this, bol_lh, PROB_MAX);
3897 inc_sp(nargs);
3898 uncommon_trap(Deoptimization::Reason_class_check,
3899 Deoptimization::Action_maybe_recompile);
3900 }
3901 layout_val = nullptr;
3902 layout_is_con = true;
3903 }
3904
3905 // Generate the initial go-slow test. Make sure we do not overflow
3906 // if length is huge (near 2Gig) or negative! We do not need
3907 // exact double-words here, just a close approximation of needed
3908 // double-words. We can't add any offset or rounding bits, lest we
3909 // take a size -1 of bytes and make it positive. Use an unsigned
3910 // compare, so negative sizes look hugely positive.
3911 int fast_size_limit = FastAllocateSizeLimit;
3912 if (layout_is_con) {
3913 assert(!StressReflectiveCode, "stress mode does not use these paths");
3914 // Increase the size limit if we have exact knowledge of array type.
3915 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
3916 assert(fast_size_limit == 0 || count_leading_zeros(fast_size_limit) > static_cast<unsigned>(LogBytesPerLong - log2_esize),
3917 "fast_size_limit (%d) overflow when shifted left by %d", fast_size_limit, LogBytesPerLong - log2_esize);
3918 fast_size_limit <<= (LogBytesPerLong - log2_esize);
3919 }
3920
3921 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
3922 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
3923
3924 // --- Size Computation ---
3925 // array_size = round_to_heap(array_header + (length << elem_shift));
3926 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
3927 // and align_to(x, y) == ((x + y-1) & ~(y-1))
3928 // The rounding mask is strength-reduced, if possible.
3929 int round_mask = MinObjAlignmentInBytes - 1;
3930 Node* header_size = nullptr;
3931 // (T_BYTE has the weakest alignment and size restrictions...)
3932 if (layout_is_con) {
3933 int hsize = Klass::layout_helper_header_size(layout_con);
3934 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3935 if ((round_mask & ~right_n_bits(eshift)) == 0)
3936 round_mask = 0; // strength-reduce it if it goes away completely
3937 assert((hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
3938 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
3939 assert(header_size_min <= hsize, "generic minimum is smallest");
3940 header_size = intcon(hsize);
3941 } else {
3942 Node* hss = intcon(Klass::_lh_header_size_shift);
3943 Node* hsm = intcon(Klass::_lh_header_size_mask);
3944 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
3945 header_size = _gvn.transform(new AndINode(header_size, hsm));
3946 }
3947
3948 Node* elem_shift = nullptr;
3949 if (layout_is_con) {
3950 int eshift = Klass::layout_helper_log2_element_size(layout_con);
3951 if (eshift != 0)
3952 elem_shift = intcon(eshift);
3953 } else {
3954 // There is no need to mask or shift this value.
3955 // The semantics of LShiftINode include an implicit mask to 0x1F.
3956 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
3957 elem_shift = layout_val;
4006 }
4007 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4008
4009 if (return_size_val != nullptr) {
4010 // This is the size
4011 (*return_size_val) = non_rounded_size;
4012 }
4013
4014 Node* size = non_rounded_size;
4015 if (round_mask != 0) {
4016 Node* mask1 = MakeConX(round_mask);
4017 size = _gvn.transform(new AddXNode(size, mask1));
4018 Node* mask2 = MakeConX(~round_mask);
4019 size = _gvn.transform(new AndXNode(size, mask2));
4020 }
4021 // else if round_mask == 0, the size computation is self-rounding
4022
4023 // Now generate allocation code
4024
4025 // The entire memory state is needed for slow path of the allocation
4026 // since GC and deoptimization can happened.
4027 Node *mem = reset_memory();
4028 set_all_memory(mem); // Create new memory state
4029
4030 if (initial_slow_test->is_Bool()) {
4031 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4032 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4033 }
4034
4035 const TypeOopPtr* ary_type = _gvn.type(klass_node)->is_klassptr()->as_instance_type();
4036 Node* valid_length_test = _gvn.intcon(1);
4037 if (ary_type->isa_aryptr()) {
4038 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4039 jint max = TypeAryPtr::max_array_length(bt);
4040 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4041 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4042 }
4043
4044 // Create the AllocateArrayNode and its result projections
4045 AllocateArrayNode* alloc
4046 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4047 control(), mem, i_o(),
4048 size, klass_node,
4049 initial_slow_test,
4050 length, valid_length_test);
4051
4052 // Cast to correct type. Note that the klass_node may be constant or not,
4053 // and in the latter case the actual array type will be inexact also.
4054 // (This happens via a non-constant argument to inline_native_newArray.)
4055 // In any case, the value of klass_node provides the desired array type.
4056 const TypeInt* length_type = _gvn.find_int_type(length);
4057 if (ary_type->isa_aryptr() && length_type != nullptr) {
4058 // Try to get a better type than POS for the size
4059 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4060 }
4061
4062 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4063
4064 array_ideal_length(alloc, ary_type, true);
4065 return javaoop;
4066 }
4067
4068 // The following "Ideal_foo" functions are placed here because they recognize
4069 // the graph shapes created by the functions immediately above.
4070
4071 //---------------------------Ideal_allocation----------------------------------
4166 void GraphKit::add_parse_predicates(int nargs) {
4167 if (ShortRunningLongLoop) {
4168 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4169 // walking up from the loop.
4170 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4171 }
4172 if (UseLoopPredicate) {
4173 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4174 if (UseProfiledLoopPredicate) {
4175 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4176 }
4177 }
4178 if (UseAutoVectorizationPredicate) {
4179 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4180 }
4181 // Loop Limit Check Predicate should be near the loop.
4182 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4183 }
4184
4185 void GraphKit::sync_kit(IdealKit& ideal) {
4186 set_all_memory(ideal.merged_memory());
4187 set_i_o(ideal.i_o());
4188 set_control(ideal.ctrl());
4189 }
4190
4191 void GraphKit::final_sync(IdealKit& ideal) {
4192 // Final sync IdealKit and graphKit.
4193 sync_kit(ideal);
4194 }
4195
4196 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4197 Node* len = load_array_length(load_String_value(str, set_ctrl));
4198 Node* coder = load_String_coder(str, set_ctrl);
4199 // Divide length by 2 if coder is UTF16
4200 return _gvn.transform(new RShiftINode(len, coder));
4201 }
4202
4203 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4204 int value_offset = java_lang_String::value_offset();
4205 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4206 false, nullptr, 0);
4207 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4208 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4209 TypeAry::make(TypeInt::BYTE, TypeInt::POS),
4210 ciTypeArrayKlass::make(T_BYTE), true, 0);
4211 Node* p = basic_plus_adr(str, str, value_offset);
4212 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4213 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4214 return load;
4215 }
4216
4217 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4218 if (!CompactStrings) {
4219 return intcon(java_lang_String::CODER_UTF16);
4220 }
4221 int coder_offset = java_lang_String::coder_offset();
4222 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4223 false, nullptr, 0);
4224 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4225
4226 Node* p = basic_plus_adr(str, str, coder_offset);
4227 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4228 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4229 return load;
4230 }
4231
4232 void GraphKit::store_String_value(Node* str, Node* value) {
4233 int value_offset = java_lang_String::value_offset();
4234 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4235 false, nullptr, 0);
4236 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4237
4238 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4239 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4240 }
4241
4242 void GraphKit::store_String_coder(Node* str, Node* value) {
4243 int coder_offset = java_lang_String::coder_offset();
4244 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4245 false, nullptr, 0);
4246 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4247
4248 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4249 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4250 }
4251
4252 // Capture src and dst memory state with a MergeMemNode
4253 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4254 if (src_type == dst_type) {
4255 // Types are equal, we don't need a MergeMemNode
4256 return memory(src_type);
4257 }
4258 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4259 record_for_igvn(merge); // fold it up later, if possible
4260 int src_idx = C->get_alias_index(src_type);
4261 int dst_idx = C->get_alias_index(dst_type);
4262 merge->set_memory_at(src_idx, memory(src_idx));
4263 merge->set_memory_at(dst_idx, memory(dst_idx));
4264 return merge;
4265 }
4338 i_char->init_req(2, AddI(i_char, intcon(2)));
4339
4340 set_control(IfFalse(iff));
4341 set_memory(st, TypeAryPtr::BYTES);
4342 }
4343
4344 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4345 if (!field->is_constant()) {
4346 return nullptr; // Field not marked as constant.
4347 }
4348 ciInstance* holder = nullptr;
4349 if (!field->is_static()) {
4350 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4351 if (const_oop != nullptr && const_oop->is_instance()) {
4352 holder = const_oop->as_instance();
4353 }
4354 }
4355 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4356 /*is_unsigned_load=*/false);
4357 if (con_type != nullptr) {
4358 return makecon(con_type);
4359 }
4360 return nullptr;
4361 }
4362
4363 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type) {
4364 const TypeOopPtr* obj_type = obj->bottom_type()->isa_oopptr();
4365 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4366 if (obj_type != nullptr && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4367 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4368 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4369 return casted_obj;
4370 }
4371 return obj;
4372 }
|
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "asm/register.hpp"
26 #include "ci/ciFlatArrayKlass.hpp"
27 #include "ci/ciInlineKlass.hpp"
28 #include "ci/ciMethod.hpp"
29 #include "ci/ciObjArray.hpp"
30 #include "ci/ciUtilities.hpp"
31 #include "classfile/javaClasses.hpp"
32 #include "compiler/compileLog.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/c2/barrierSetC2.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/flatArrayKlass.hpp"
38 #include "opto/addnode.hpp"
39 #include "opto/callnode.hpp"
40 #include "opto/castnode.hpp"
41 #include "opto/convertnode.hpp"
42 #include "opto/graphKit.hpp"
43 #include "opto/idealKit.hpp"
44 #include "opto/inlinetypenode.hpp"
45 #include "opto/intrinsicnode.hpp"
46 #include "opto/locknode.hpp"
47 #include "opto/machnode.hpp"
48 #include "opto/memnode.hpp"
49 #include "opto/multnode.hpp"
50 #include "opto/narrowptrnode.hpp"
51 #include "opto/opaquenode.hpp"
52 #include "opto/parse.hpp"
53 #include "opto/reachability.hpp"
54 #include "opto/rootnode.hpp"
55 #include "opto/runtime.hpp"
56 #include "opto/subtypenode.hpp"
57 #include "runtime/arguments.hpp"
58 #include "runtime/deoptimization.hpp"
59 #include "runtime/sharedRuntime.hpp"
60 #include "runtime/stubRoutines.hpp"
61 #include "utilities/bitMap.inline.hpp"
62 #include "utilities/growableArray.hpp"
63 #include "utilities/powerOfTwo.hpp"
64
65 //----------------------------GraphKit-----------------------------------------
66 // Main utility constructor.
67 GraphKit::GraphKit(JVMState* jvms, PhaseGVN* gvn)
68 : Phase(Phase::Parser),
69 _env(C->env()),
70 _gvn((gvn != nullptr) ? *gvn : *C->initial_gvn()),
71 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
72 {
73 assert(gvn == nullptr || !gvn->is_IterGVN() || gvn->is_IterGVN()->delay_transform(), "delay transform should be enabled");
74 _exceptions = jvms->map()->next_exception();
75 if (_exceptions != nullptr) jvms->map()->set_next_exception(nullptr);
76 set_jvms(jvms);
77 #ifdef ASSERT
78 if (_gvn.is_IterGVN() != nullptr) {
79 assert(_gvn.is_IterGVN()->delay_transform(), "Transformation must be delayed if IterGVN is used");
80 // Save the initial size of _for_igvn worklist for verification (see ~GraphKit)
81 _worklist_size = _gvn.C->igvn_worklist()->size();
82 }
83 #endif
84 }
85
86 // Private constructor for parser.
87 GraphKit::GraphKit()
88 : Phase(Phase::Parser),
89 _env(C->env()),
90 _gvn(*C->initial_gvn()),
91 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2())
92 {
93 _exceptions = nullptr;
94 set_map(nullptr);
95 DEBUG_ONLY(_sp = -99);
96 DEBUG_ONLY(set_bci(-99));
97 }
98
99 GraphKit::GraphKit(const SafePointNode* sft, PhaseIterGVN& igvn)
100 : Phase(Phase::Parser),
101 _env(C->env()),
102 _gvn(igvn),
103 _exceptions(nullptr),
104 _barrier_set(BarrierSet::barrier_set()->barrier_set_c2()) {
105 assert(igvn.delay_transform(), "must delay transformation during macro expansion");
106 assert(sft->next_exception() == nullptr, "must not have a pending exception");
107 JVMState* cloned_jvms = sft->jvms()->clone_deep(C);
108 SafePointNode* cloned_map = new SafePointNode(sft->req(), cloned_jvms);
109 for (uint i = 0; i < sft->req(); i++) {
110 cloned_map->init_req(i, sft->in(i));
111 }
112 igvn.record_for_igvn(cloned_map);
113 for (JVMState* current = cloned_jvms; current != nullptr; current = current->caller()) {
114 current->set_map(cloned_map);
115 }
116 set_jvms(cloned_jvms);
117 set_all_memory(reset_memory());
118 }
119
120 //---------------------------clean_stack---------------------------------------
121 // Clear away rubbish from the stack area of the JVM state.
122 // This destroys any arguments that may be waiting on the stack.
123 void GraphKit::clean_stack(int from_sp) {
124 SafePointNode* map = this->map();
125 JVMState* jvms = this->jvms();
126 int stk_size = jvms->stk_size();
127 int stkoff = jvms->stkoff();
128 Node* top = this->top();
129 for (int i = from_sp; i < stk_size; i++) {
130 if (map->in(stkoff + i) != top) {
131 map->set_req(stkoff + i, top);
132 }
133 }
134 }
135
136
137 //--------------------------------sync_jvms-----------------------------------
138 // Make sure our current jvms agrees with our parse state.
367 }
368 static inline void add_one_req(Node* dstphi, Node* src) {
369 assert(is_hidden_merge(dstphi), "must be a special merge node");
370 assert(!is_hidden_merge(src), "must not be a special merge node");
371 dstphi->add_req(src);
372 }
373
374 //-----------------------combine_exception_states------------------------------
375 // This helper function combines exception states by building phis on a
376 // specially marked state-merging region. These regions and phis are
377 // untransformed, and can build up gradually. The region is marked by
378 // having a control input of its exception map, rather than null. Such
379 // regions do not appear except in this function, and in use_exception_state.
380 void GraphKit::combine_exception_states(SafePointNode* ex_map, SafePointNode* phi_map) {
381 if (failing_internal()) {
382 return; // dying anyway...
383 }
384 JVMState* ex_jvms = ex_map->_jvms;
385 assert(ex_jvms->same_calls_as(phi_map->_jvms), "consistent call chains");
386 assert(ex_jvms->stkoff() == phi_map->_jvms->stkoff(), "matching locals");
387 // TODO 8325632 Re-enable
388 // assert(ex_jvms->sp() == phi_map->_jvms->sp(), "matching stack sizes");
389 assert(ex_jvms->monoff() == phi_map->_jvms->monoff(), "matching JVMS");
390 assert(ex_jvms->scloff() == phi_map->_jvms->scloff(), "matching scalar replaced objects");
391 assert(ex_map->req() == phi_map->req(), "matching maps");
392 uint tos = ex_jvms->stkoff() + ex_jvms->sp();
393 Node* hidden_merge_mark = root();
394 Node* region = phi_map->control();
395 MergeMemNode* phi_mem = phi_map->merged_memory();
396 MergeMemNode* ex_mem = ex_map->merged_memory();
397 if (region->in(0) != hidden_merge_mark) {
398 // The control input is not (yet) a specially-marked region in phi_map.
399 // Make it so, and build some phis.
400 region = new RegionNode(2);
401 _gvn.set_type(region, Type::CONTROL);
402 region->set_req(0, hidden_merge_mark); // marks an internal ex-state
403 region->init_req(1, phi_map->control());
404 phi_map->set_control(region);
405 Node* io_phi = PhiNode::make(region, phi_map->i_o(), Type::ABIO);
406 record_for_igvn(io_phi);
407 _gvn.set_type(io_phi, Type::ABIO);
408 phi_map->set_i_o(io_phi);
938 if (PrintMiscellaneous && (Verbose || WizardMode)) {
939 tty->print_cr("Zombie local %d: ", local);
940 jvms->dump();
941 }
942 return false;
943 }
944 }
945 }
946 return true;
947 }
948
949 #endif //ASSERT
950
951 // Helper function for enforcing certain bytecodes to reexecute if deoptimization happens.
952 static bool should_reexecute_implied_by_bytecode(JVMState *jvms, bool is_anewarray) {
953 ciMethod* cur_method = jvms->method();
954 int cur_bci = jvms->bci();
955 if (cur_method != nullptr && cur_bci != InvocationEntryBci) {
956 Bytecodes::Code code = cur_method->java_code_at_bci(cur_bci);
957 return Interpreter::bytecode_should_reexecute(code) ||
958 (is_anewarray && (code == Bytecodes::_multianewarray));
959 // Reexecute _multianewarray bytecode which was replaced with
960 // sequence of [a]newarray. See Parse::do_multianewarray().
961 //
962 // Note: interpreter should not have it set since this optimization
963 // is limited by dimensions and guarded by flag so in some cases
964 // multianewarray() runtime calls will be generated and
965 // the bytecode should not be reexecutes (stack will not be reset).
966 } else {
967 return false;
968 }
969 }
970
971 // Helper function for adding JVMState and debug information to node
972 void GraphKit::add_safepoint_edges(SafePointNode* call, bool must_throw) {
973 // Add the safepoint edges to the call (or other safepoint).
974
975 // Make sure dead locals are set to top. This
976 // should help register allocation time and cut down on the size
977 // of the deoptimization information.
978 assert(dead_locals_are_killed(), "garbage in debug info before safepoint");
1054 uint p = debug_start; // walks forward in [debug_start, debug_end)
1055 uint j, k, l;
1056 SafePointNode* in_map = in_jvms->map();
1057 out_jvms->set_map(call);
1058
1059 if (can_prune_locals) {
1060 assert(in_jvms->method() == out_jvms->method(), "sanity");
1061 // If the current throw can reach an exception handler in this JVMS,
1062 // then we must keep everything live that can reach that handler.
1063 // As a quick and dirty approximation, we look for any handlers at all.
1064 if (in_jvms->method()->has_exception_handlers()) {
1065 can_prune_locals = false;
1066 }
1067 }
1068
1069 // Add the Locals
1070 k = in_jvms->locoff();
1071 l = in_jvms->loc_size();
1072 out_jvms->set_locoff(p);
1073 if (!can_prune_locals) {
1074 for (j = 0; j < l; j++) {
1075 call->set_req(p++, in_map->in(k + j));
1076 }
1077 } else {
1078 p += l; // already set to top above by add_req_batch
1079 }
1080
1081 // Add the Expression Stack
1082 k = in_jvms->stkoff();
1083 l = in_jvms->sp();
1084 out_jvms->set_stkoff(p);
1085 if (!can_prune_locals) {
1086 for (j = 0; j < l; j++) {
1087 call->set_req(p++, in_map->in(k + j));
1088 }
1089 } else if (can_prune_locals && stack_slots_not_pruned != 0) {
1090 // Divide stack into {S0,...,S1}, where S0 is set to top.
1091 uint s1 = stack_slots_not_pruned;
1092 stack_slots_not_pruned = 0; // for next iteration
1093 if (s1 > l) s1 = l;
1094 uint s0 = l - s1;
1095 p += s0; // skip the tops preinstalled by add_req_batch
1096 for (j = s0; j < l; j++)
1097 call->set_req(p++, in_map->in(k+j));
1098 } else {
1099 p += l; // already set to top above by add_req_batch
1100 }
1101
1102 // Add the Monitors
1103 k = in_jvms->monoff();
1104 l = in_jvms->mon_size();
1105 out_jvms->set_monoff(p);
1106 for (j = 0; j < l; j++)
1107 call->set_req(p++, in_map->in(k+j));
1108
1296 Node* conv = _gvn.transform( new ConvI2LNode(offset));
1297 Node* mask = _gvn.transform(ConLNode::make((julong) max_juint));
1298 return _gvn.transform( new AndLNode(conv, mask) );
1299 }
1300
1301 Node* GraphKit::ConvL2I(Node* offset) {
1302 // short-circuit a common case
1303 jlong offset_con = find_long_con(offset, (jlong)Type::OffsetBot);
1304 if (offset_con != (jlong)Type::OffsetBot) {
1305 return intcon((int) offset_con);
1306 }
1307 return _gvn.transform( new ConvL2INode(offset));
1308 }
1309
1310 //-------------------------load_object_klass-----------------------------------
1311 Node* GraphKit::load_object_klass(Node* obj) {
1312 // Special-case a fresh allocation to avoid building nodes:
1313 Node* akls = AllocateNode::Ideal_klass(obj, &_gvn);
1314 if (akls != nullptr) return akls;
1315 Node* k_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
1316 return _gvn.transform(LoadKlassNode::make(_gvn, immutable_memory(), k_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
1317 }
1318
1319 //-------------------------load_array_length-----------------------------------
1320 Node* GraphKit::load_array_length(Node* array) {
1321 // Special-case a fresh allocation to avoid building nodes:
1322 AllocateArrayNode* alloc = AllocateArrayNode::Ideal_array_allocation(array);
1323 Node *alen;
1324 if (alloc == nullptr) {
1325 Node *r_adr = basic_plus_adr(array, arrayOopDesc::length_offset_in_bytes());
1326 alen = _gvn.transform( new LoadRangeNode(nullptr, immutable_memory(), r_adr, TypeInt::POS));
1327 } else {
1328 alen = array_ideal_length(alloc, _gvn.type(array)->is_oopptr(), false);
1329 }
1330 return alen;
1331 }
1332
1333 Node* GraphKit::array_ideal_length(AllocateArrayNode* alloc,
1334 const TypeOopPtr* oop_type,
1335 bool replace_length_in_map) {
1336 Node* length = alloc->Ideal_length();
1345 replace_in_map(length, ccast);
1346 }
1347 return ccast;
1348 }
1349 }
1350 return length;
1351 }
1352
1353 //------------------------------do_null_check----------------------------------
1354 // Helper function to do a null pointer check. Returned value is
1355 // the incoming address with null casted away. You are allowed to use the
1356 // not-null value only if you are control dependent on the test.
1357 #ifndef PRODUCT
1358 extern uint explicit_null_checks_inserted,
1359 explicit_null_checks_elided;
1360 #endif
1361 Node* GraphKit::null_check_common(Node* value, BasicType type,
1362 // optional arguments for variations:
1363 bool assert_null,
1364 Node* *null_control,
1365 bool speculative,
1366 bool null_marker_check) {
1367 assert(!assert_null || null_control == nullptr, "not both at once");
1368 if (stopped()) return top();
1369 NOT_PRODUCT(explicit_null_checks_inserted++);
1370
1371 if (value->is_InlineType()) {
1372 // Null checking a scalarized but nullable inline type. Check the null marker
1373 // input instead of the oop input to avoid keeping buffer allocations alive.
1374 InlineTypeNode* vtptr = value->as_InlineType();
1375 while (vtptr->get_oop()->is_InlineType()) {
1376 vtptr = vtptr->get_oop()->as_InlineType();
1377 }
1378 null_check_common(vtptr->get_null_marker(), T_INT, assert_null, null_control, speculative, true);
1379 if (stopped()) {
1380 return top();
1381 }
1382 if (assert_null) {
1383 // TODO 8350865 Scalarize here (this leads to failures with TestLWorld::test45)
1384 // vtptr = InlineTypeNode::make_null(_gvn, vtptr->type()->inline_klass());
1385 // replace_in_map(value, vtptr);
1386 // return vtptr;
1387 replace_in_map(value, null());
1388 return null();
1389 }
1390 bool do_replace_in_map = (null_control == nullptr || (*null_control) == top());
1391 return cast_not_null(value, do_replace_in_map);
1392 }
1393
1394 // Construct null check
1395 Node *chk = nullptr;
1396 switch(type) {
1397 case T_LONG : chk = new CmpLNode(value, _gvn.zerocon(T_LONG)); break;
1398 case T_INT : chk = new CmpINode(value, _gvn.intcon(0)); break;
1399 case T_ARRAY : // fall through
1400 type = T_OBJECT; // simplify further tests
1401 case T_OBJECT : {
1402 const Type *t = _gvn.type( value );
1403
1404 const TypeOopPtr* tp = t->isa_oopptr();
1405 if (tp != nullptr && !tp->is_loaded()
1406 // Only for do_null_check, not any of its siblings:
1407 && !assert_null && null_control == nullptr) {
1408 // Usually, any field access or invocation on an unloaded oop type
1409 // will simply fail to link, since the statically linked class is
1410 // likely also to be unloaded. However, in -Xcomp mode, sometimes
1411 // the static class is loaded but the sharper oop type is not.
1412 // Rather than checking for this obscure case in lots of places,
1413 // we simply observe that a null check on an unloaded class
1477 }
1478 Node *oldcontrol = control();
1479 set_control(cfg);
1480 Node *res = cast_not_null(value);
1481 set_control(oldcontrol);
1482 NOT_PRODUCT(explicit_null_checks_elided++);
1483 return res;
1484 }
1485 cfg = IfNode::up_one_dom(cfg, /*linear_only=*/ true);
1486 if (cfg == nullptr) break; // Quit at region nodes
1487 depth++;
1488 }
1489 }
1490
1491 //-----------
1492 // Branch to failure if null
1493 float ok_prob = PROB_MAX; // a priori estimate: nulls never happen
1494 Deoptimization::DeoptReason reason;
1495 if (assert_null) {
1496 reason = Deoptimization::reason_null_assert(speculative);
1497 } else if (type == T_OBJECT || null_marker_check) {
1498 reason = Deoptimization::reason_null_check(speculative);
1499 } else {
1500 reason = Deoptimization::Reason_div0_check;
1501 }
1502 // %%% Since Reason_unhandled is not recorded on a per-bytecode basis,
1503 // ciMethodData::has_trap_at will return a conservative -1 if any
1504 // must-be-null assertion has failed. This could cause performance
1505 // problems for a method after its first do_null_assert failure.
1506 // Consider using 'Reason_class_check' instead?
1507
1508 // To cause an implicit null check, we set the not-null probability
1509 // to the maximum (PROB_MAX). For an explicit check the probability
1510 // is set to a smaller value.
1511 if (null_control != nullptr || too_many_traps(reason)) {
1512 // probability is less likely
1513 ok_prob = PROB_LIKELY_MAG(3);
1514 } else if (!assert_null &&
1515 (ImplicitNullCheckThreshold > 0) &&
1516 method() != nullptr &&
1517 (method()->method_data()->trap_count(reason)
1551 }
1552
1553 if (assert_null) {
1554 // Cast obj to null on this path.
1555 replace_in_map(value, zerocon(type));
1556 return zerocon(type);
1557 }
1558
1559 // Cast obj to not-null on this path, if there is no null_control.
1560 // (If there is a null_control, a non-null value may come back to haunt us.)
1561 if (type == T_OBJECT) {
1562 Node* cast = cast_not_null(value, false);
1563 if (null_control == nullptr || (*null_control) == top())
1564 replace_in_map(value, cast);
1565 value = cast;
1566 }
1567
1568 return value;
1569 }
1570
1571 //------------------------------cast_not_null----------------------------------
1572 // Cast obj to not-null on this path
1573 Node* GraphKit::cast_not_null(Node* obj, bool do_replace_in_map) {
1574 if (obj->is_InlineType()) {
1575 Node* vt = obj->isa_InlineType()->clone_if_required(&gvn(), map(), do_replace_in_map);
1576 vt->as_InlineType()->set_null_marker(_gvn);
1577 vt = _gvn.transform(vt);
1578 if (do_replace_in_map) {
1579 replace_in_map(obj, vt);
1580 }
1581 return vt;
1582 }
1583 const Type *t = _gvn.type(obj);
1584 const Type *t_not_null = t->join_speculative(TypePtr::NOTNULL);
1585 // Object is already not-null?
1586 if( t == t_not_null ) return obj;
1587
1588 Node* cast = new CastPPNode(control(), obj,t_not_null);
1589 cast = _gvn.transform( cast );
1590
1591 // Scan for instances of 'obj' in the current JVM mapping.
1592 // These instances are known to be not-null after the test.
1593 if (do_replace_in_map)
1594 replace_in_map(obj, cast);
1595
1596 return cast; // Return casted value
1597 }
1598
1599 // Sometimes in intrinsics, we implicitly know an object is not null
1600 // (there's no actual null check) so we can cast it to not null. In
1601 // the course of optimizations, the input to the cast can become null.
1602 // In that case that data path will die and we need the control path
1657 Node* GraphKit::memory(uint alias_idx) {
1658 MergeMemNode* mem = merged_memory();
1659 Node* p = mem->memory_at(alias_idx);
1660 assert(p != mem->empty_memory(), "empty");
1661 _gvn.set_type(p, Type::MEMORY); // must be mapped
1662 return p;
1663 }
1664
1665 //-----------------------------reset_memory------------------------------------
1666 Node* GraphKit::reset_memory() {
1667 Node* mem = map()->memory();
1668 // do not use this node for any more parsing!
1669 DEBUG_ONLY( map()->set_memory((Node*)nullptr) );
1670 return _gvn.transform( mem );
1671 }
1672
1673 //------------------------------set_all_memory---------------------------------
1674 void GraphKit::set_all_memory(Node* newmem) {
1675 Node* mergemem = MergeMemNode::make(newmem);
1676 gvn().set_type_bottom(mergemem);
1677 if (_gvn.is_IterGVN() != nullptr) {
1678 record_for_igvn(mergemem);
1679 }
1680 map()->set_memory(mergemem);
1681 }
1682
1683 //------------------------------set_all_memory_call----------------------------
1684 void GraphKit::set_all_memory_call(Node* call, bool separate_io_proj) {
1685 Node* newmem = _gvn.transform( new ProjNode(call, TypeFunc::Memory, separate_io_proj) );
1686 set_all_memory(newmem);
1687 }
1688
1689 //=============================================================================
1690 //
1691 // parser factory methods for MemNodes
1692 //
1693 // These are layered on top of the factory methods in LoadNode and StoreNode,
1694 // and integrate with the parser's memory state and _gvn engine.
1695 //
1696
1697 // factory methods in "int adr_idx"
1698 Node* GraphKit::make_load(Node* ctl, Node* adr, const Type* t, BasicType bt,
1699 MemNode::MemOrd mo,
1700 LoadNode::ControlDependency control_dependency,
1701 bool require_atomic_access,
1702 bool unaligned,
1703 bool mismatched,
1704 bool unsafe,
1705 uint8_t barrier_data) {
1706 int adr_idx = C->get_alias_index(_gvn.type(adr)->isa_ptr());
1707 assert(adr_idx != Compile::AliasIdxTop, "use other make_load factory" );
1708 const TypePtr* adr_type = nullptr; // debug-mode-only argument
1709 DEBUG_ONLY(adr_type = C->get_adr_type(adr_idx));
1710 Node* mem = memory(adr_idx);
1711 Node* ld = LoadNode::make(_gvn, ctl, mem, adr, adr_type, t, bt, mo, control_dependency, require_atomic_access, unaligned, mismatched, unsafe, barrier_data);
1712 ld = _gvn.transform(ld);
1713
1714 if (((bt == T_OBJECT) && C->do_escape_analysis()) || C->eliminate_boxing()) {
1715 // Improve graph before escape analysis and boxing elimination.
1716 record_for_igvn(ld);
1717 if (ld->is_DecodeN()) {
1718 // Also record the actual load (LoadN) in case ld is DecodeN. In some
1719 // rare corner cases, ld->in(1) can be something other than LoadN (e.g.,
1720 // a Phi). Recording such cases is still perfectly sound, but may be
1721 // unnecessary and result in some minor IGVN overhead.
1722 record_for_igvn(ld->in(1));
1723 }
1724 }
1725 return ld;
1726 }
1727
1728 Node* GraphKit::store_to_memory(Node* ctl, Node* adr, Node *val, BasicType bt,
1729 MemNode::MemOrd mo,
1730 bool require_atomic_access,
1731 bool unaligned,
1732 bool mismatched,
1733 bool unsafe,
1747 if (unsafe) {
1748 st->as_Store()->set_unsafe_access();
1749 }
1750 st->as_Store()->set_barrier_data(barrier_data);
1751 st = _gvn.transform(st);
1752 set_memory(st, adr_idx);
1753 // Back-to-back stores can only remove intermediate store with DU info
1754 // so push on worklist for optimizer.
1755 if (mem->req() > MemNode::Address && adr == mem->in(MemNode::Address))
1756 record_for_igvn(st);
1757
1758 return st;
1759 }
1760
1761 Node* GraphKit::access_store_at(Node* obj,
1762 Node* adr,
1763 const TypePtr* adr_type,
1764 Node* val,
1765 const Type* val_type,
1766 BasicType bt,
1767 DecoratorSet decorators,
1768 bool safe_for_replace,
1769 const InlineTypeNode* vt) {
1770 // Transformation of a value which could be null pointer (CastPP #null)
1771 // could be delayed during Parse (for example, in adjust_map_after_if()).
1772 // Execute transformation here to avoid barrier generation in such case.
1773 if (_gvn.type(val) == TypePtr::NULL_PTR) {
1774 val = _gvn.makecon(TypePtr::NULL_PTR);
1775 }
1776
1777 if (stopped()) {
1778 return top(); // Dead path ?
1779 }
1780
1781 assert(val != nullptr, "not dead path");
1782 if (val->is_InlineType()) {
1783 // Store to non-flat field. Buffer the inline type and make sure
1784 // the store is re-executed if the allocation triggers deoptimization.
1785 PreserveReexecuteState preexecs(this);
1786 jvms()->set_should_reexecute(true);
1787 val = val->as_InlineType()->buffer(this, safe_for_replace);
1788 }
1789
1790 C2AccessValuePtr addr(adr, adr_type);
1791 C2AccessValue value(val, val_type);
1792 C2ParseAccess access(this, decorators | C2_WRITE_ACCESS, bt, obj, addr, nullptr, vt);
1793 if (access.is_raw()) {
1794 return _barrier_set->BarrierSetC2::store_at(access, value);
1795 } else {
1796 return _barrier_set->store_at(access, value);
1797 }
1798 }
1799
1800 Node* GraphKit::access_load_at(Node* obj, // containing obj
1801 Node* adr, // actual address to store val at
1802 const TypePtr* adr_type,
1803 const Type* val_type,
1804 BasicType bt,
1805 DecoratorSet decorators,
1806 Node* ctl) {
1807 if (stopped()) {
1808 return top(); // Dead path ?
1809 }
1810
1811 SavedState old_state(this);
1812 C2AccessValuePtr addr(adr, adr_type);
1813 C2ParseAccess access(this, decorators | C2_READ_ACCESS, bt, obj, addr, ctl);
1814 Node* load;
1815 if (access.is_raw()) {
1816 load = _barrier_set->BarrierSetC2::load_at(access, val_type);
1817 } else {
1818 load = _barrier_set->load_at(access, val_type);
1819 }
1820
1821 // Restore the previous state only if the load got folded to a constant
1822 // and we can discard any barriers that might have been added.
1823 if (load == nullptr || !load->is_Con()) {
1824 old_state.discard();
1825 }
1826 return load;
1827 }
1828
1829 Node* GraphKit::access_load(Node* adr, // actual address to load val at
1830 const Type* val_type,
1831 BasicType bt,
1832 DecoratorSet decorators) {
1833 if (stopped()) {
1915 Node* new_val,
1916 const Type* value_type,
1917 BasicType bt,
1918 DecoratorSet decorators) {
1919 C2AccessValuePtr addr(adr, adr_type);
1920 C2AtomicParseAccess access(this, decorators | C2_READ_ACCESS | C2_WRITE_ACCESS, bt, obj, addr, alias_idx);
1921 if (access.is_raw()) {
1922 return _barrier_set->BarrierSetC2::atomic_add_at(access, new_val, value_type);
1923 } else {
1924 return _barrier_set->atomic_add_at(access, new_val, value_type);
1925 }
1926 }
1927
1928 void GraphKit::access_clone(Node* src, Node* dst, Node* size, bool is_array) {
1929 return _barrier_set->clone(this, src, dst, size, is_array);
1930 }
1931
1932 //-------------------------array_element_address-------------------------
1933 Node* GraphKit::array_element_address(Node* ary, Node* idx, BasicType elembt,
1934 const TypeInt* sizetype, Node* ctrl) {
1935 const TypeAryPtr* arytype = _gvn.type(ary)->is_aryptr();
1936 uint shift;
1937 uint header;
1938 if (arytype->is_flat() && arytype->klass_is_exact()) {
1939 // We can only determine the flat array layout statically if the klass is exact. Otherwise, we could have different
1940 // value classes at runtime with a potentially different layout. The caller needs to fall back to call
1941 // load/store_unknown_inline_Type() at runtime. We could return a sentinel node for the non-exact case but that
1942 // might mess with other GVN transformations in between. Thus, we just continue in the else branch normally, even
1943 // though we don't need the address node in this case and throw it away again.
1944 shift = arytype->flat_log_elem_size();
1945 header = arrayOopDesc::base_offset_in_bytes(T_FLAT_ELEMENT);
1946 } else {
1947 shift = exact_log2(type2aelembytes(elembt));
1948 header = arrayOopDesc::base_offset_in_bytes(elembt);
1949 }
1950
1951 // short-circuit a common case (saves lots of confusing waste motion)
1952 jint idx_con = find_int_con(idx, -1);
1953 if (idx_con >= 0) {
1954 intptr_t offset = header + ((intptr_t)idx_con << shift);
1955 return basic_plus_adr(ary, offset);
1956 }
1957
1958 // must be correct type for alignment purposes
1959 Node* base = basic_plus_adr(ary, header);
1960 idx = Compile::conv_I2X_index(&_gvn, idx, sizetype, ctrl);
1961 Node* scale = _gvn.transform( new LShiftXNode(idx, intcon(shift)) );
1962 return basic_plus_adr(ary, base, scale);
1963 }
1964
1965 Node* GraphKit::cast_to_flat_array(Node* array, ciInlineKlass* elem_vk) {
1966 assert(elem_vk->maybe_flat_in_array(), "no flat array for %s", elem_vk->name()->as_utf8());
1967 if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_nullable_atomic_layout()) {
1968 return cast_to_flat_array_exact(array, elem_vk, true, false);
1969 } else if (!elem_vk->has_nullable_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1970 return cast_to_flat_array_exact(array, elem_vk, true, true);
1971 } else if (!elem_vk->has_null_free_atomic_layout() && !elem_vk->has_null_free_non_atomic_layout()) {
1972 return cast_to_flat_array_exact(array, elem_vk, false, true);
1973 }
1974
1975 bool is_null_free = false;
1976 if (!elem_vk->has_nullable_atomic_layout()) {
1977 // Element does not have a nullable flat layout, cannot be nullable
1978 is_null_free = true;
1979 }
1980
1981 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, false);
1982 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1983 arytype = arytype->cast_to_flat(true)->cast_to_null_free(is_null_free);
1984 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1985 }
1986
1987 Node* GraphKit::cast_to_flat_array_exact(Node* array, ciInlineKlass* elem_vk, bool is_null_free, bool is_atomic) {
1988 assert(is_null_free || is_atomic, "nullable arrays must be atomic");
1989 ciArrayKlass* array_klass = ciObjArrayKlass::make(elem_vk, true, is_null_free, is_atomic);
1990 const TypeAryPtr* arytype = TypeOopPtr::make_from_klass(array_klass)->isa_aryptr();
1991 assert(arytype->klass_is_exact(), "inconsistency");
1992 assert(arytype->is_flat(), "inconsistency");
1993 assert(arytype->is_null_free() == is_null_free, "inconsistency");
1994 assert(arytype->is_not_null_free() == !is_null_free, "inconsistency");
1995 return _gvn.transform(new CheckCastPPNode(control(), array, arytype, ConstraintCastNode::DependencyType::NonFloatingNarrowing));
1996 }
1997
1998 //-------------------------load_array_element-------------------------
1999 Node* GraphKit::load_array_element(Node* ary, Node* idx, const TypeAryPtr* arytype, bool set_ctrl) {
2000 const Type* elemtype = arytype->elem();
2001 BasicType elembt = elemtype->array_element_basic_type();
2002 Node* adr = array_element_address(ary, idx, elembt, arytype->size());
2003 if (elembt == T_NARROWOOP) {
2004 elembt = T_OBJECT; // To satisfy switch in LoadNode::make()
2005 }
2006 Node* ld = access_load_at(ary, adr, arytype, elemtype, elembt,
2007 IN_HEAP | IS_ARRAY | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0));
2008 return ld;
2009 }
2010
2011 //-------------------------set_arguments_for_java_call-------------------------
2012 // Arguments (pre-popped from the stack) are taken from the JVMS.
2013 void GraphKit::set_arguments_for_java_call(CallJavaNode* call) {
2014 PreserveReexecuteState preexecs(this);
2015 if (Arguments::is_valhalla_enabled()) {
2016 // Make sure the call is "re-executed", if buffering of inline type arguments triggers deoptimization.
2017 // At this point, the call hasn't been executed yet, so we will only ever execute the call once.
2018 jvms()->set_should_reexecute(true);
2019 int arg_size = method()->get_declared_signature_at_bci(bci())->arg_size_for_bc(java_bc());
2020 inc_sp(arg_size);
2021 }
2022 // Add the call arguments
2023 const TypeTuple* domain = call->tf()->domain_sig();
2024 uint nargs = domain->cnt();
2025 int arg_num = 0;
2026 for (uint i = TypeFunc::Parms, idx = TypeFunc::Parms; i < nargs; i++) {
2027 uint arg_idx = i - TypeFunc::Parms;
2028 Node* arg = argument(arg_idx);
2029 const Type* t = domain->field_at(i);
2030 if (t->is_inlinetypeptr() && !call->method()->mismatch() && call->method()->is_scalarized_arg(arg_num)) {
2031 // We don't pass inline type arguments by reference but instead pass each field of the inline type
2032 if (!arg->is_InlineType()) {
2033 // There are 2 cases in which the argument has not been scalarized
2034 if (_gvn.type(arg)->is_zero_type()) {
2035 arg = InlineTypeNode::make_null(_gvn, t->inline_klass());
2036 } else {
2037 // During parsing, a method is called with an abstract (or j.l.Object) receiver, the
2038 // receiver is a non-scalarized oop. CHA or IGVN might then prove that the receiver
2039 // type must be an exact value class. The method is devirtualized, and replaced with
2040 // a direct call with a scalarized receiver instead.
2041 assert(arg_idx == 0 && !call->method()->is_static(), "must be the receiver");
2042 assert(call->is_optimized_virtual(), "must be during devirtualization of calls");
2043 arg = InlineTypeNode::make_from_oop(this, arg, t->inline_klass());
2044 }
2045 }
2046 InlineTypeNode* vt = arg->as_InlineType();
2047 vt->pass_fields(this, call, idx, true, !t->maybe_null(), true);
2048 // If an inline type argument is passed as fields, attach the Method* to the call site
2049 // to be able to access the extended signature later via attached_method_before_pc().
2050 // For example, see CompiledMethod::preserve_callee_argument_oops().
2051 call->set_override_symbolic_info(true);
2052 // Register a calling convention dependency on the callee method to make sure that this method is deoptimized and
2053 // re-compiled with a non-scalarized calling convention if the callee method is later marked as mismatched.
2054 C->dependencies()->assert_mismatch_calling_convention(call->method());
2055 arg_num++;
2056 continue;
2057 } else if (arg->is_InlineType()) {
2058 // Pass inline type argument via oop to callee
2059 arg = arg->as_InlineType()->buffer(this, true);
2060 }
2061 if (t != Type::HALF) {
2062 arg_num++;
2063 }
2064 call->init_req(idx++, arg);
2065 }
2066 }
2067
2068 //---------------------------set_edges_for_java_call---------------------------
2069 // Connect a newly created call into the current JVMS.
2070 // A return value node (if any) is returned from set_edges_for_java_call.
2071 void GraphKit::set_edges_for_java_call(CallJavaNode* call, bool must_throw, bool separate_io_proj) {
2072
2073 // Add the predefined inputs:
2074 call->init_req( TypeFunc::Control, control() );
2075 call->init_req( TypeFunc::I_O , i_o() );
2076 call->init_req( TypeFunc::Memory , reset_memory() );
2077 call->init_req( TypeFunc::FramePtr, frameptr() );
2078 call->init_req( TypeFunc::ReturnAdr, top() );
2079
2080 add_safepoint_edges(call, must_throw);
2081
2082 Node* xcall = _gvn.transform(call);
2083
2084 if (xcall == top()) {
2085 set_control(top());
2086 return;
2087 }
2088 assert(xcall == call, "call identity is stable");
2089
2090 // Re-use the current map to produce the result.
2091
2092 set_control(_gvn.transform(new ProjNode(call, TypeFunc::Control)));
2093 set_i_o( _gvn.transform(new ProjNode(call, TypeFunc::I_O , separate_io_proj)));
2094 set_all_memory_call(xcall, separate_io_proj);
2095
2096 //return xcall; // no need, caller already has it
2097 }
2098
2099 Node* GraphKit::set_results_for_java_call(CallJavaNode* call, bool separate_io_proj, bool deoptimize) {
2100 if (stopped()) return top(); // maybe the call folded up?
2101
2102 // Note: Since any out-of-line call can produce an exception,
2103 // we always insert an I_O projection from the call into the result.
2104
2105 make_slow_call_ex(call, env()->Throwable_klass(), separate_io_proj, deoptimize);
2106
2107 if (separate_io_proj) {
2108 // The caller requested separate projections be used by the fall
2109 // through and exceptional paths, so replace the projections for
2110 // the fall through path.
2111 set_i_o(_gvn.transform( new ProjNode(call, TypeFunc::I_O) ));
2112 set_all_memory(_gvn.transform( new ProjNode(call, TypeFunc::Memory) ));
2113 }
2114
2115 // Capture the return value, if any.
2116 Node* ret;
2117 if (call->method() == nullptr || call->method()->return_type()->basic_type() == T_VOID) {
2118 ret = top();
2119 } else if (call->tf()->returns_inline_type_as_fields()) {
2120 // Return of multiple values (inline type fields): we create a
2121 // InlineType node, each field is a projection from the call.
2122 ciInlineKlass* vk = call->method()->return_type()->as_inline_klass();
2123 uint base_input = TypeFunc::Parms;
2124 ret = InlineTypeNode::make_from_multi(this, call, vk, base_input, false, false);
2125 // If we run out of registers to store the null marker, we need to reserve an extra
2126 // slot to store it on the stack. Unfortunately, we only know if stack slot is needed
2127 // when matching the call (see Matcher::return_values_mask), so we are conservative here.
2128 C->set_needs_nm_slot(true);
2129 } else {
2130 ret = _gvn.transform(new ProjNode(call, TypeFunc::Parms));
2131 ciType* t = call->method()->return_type();
2132 if (!t->is_loaded() && InlineTypeReturnedAsFields) {
2133 // The return type is unloaded but the callee might later be C2 compiled and then return
2134 // in scalarized form when the return type is loaded. Handle this similar to what we do in
2135 // PhaseMacroExpand::expand_mh_intrinsic_return by calling into the runtime to buffer.
2136 // It's a bit unfortunate because we will deopt anyway but the interpreter needs an oop.
2137 IdealKit ideal(this);
2138 IdealVariable res(ideal);
2139 ideal.declarations_done();
2140 // Change return type of call to scalarized return
2141 const TypeFunc* tf = call->_tf;
2142 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2143 const TypeFunc* new_tf = TypeFunc::make(tf->domain_sig(), tf->domain_cc(), tf->range_sig(), domain, true);
2144 call->_tf = new_tf;
2145 _gvn.set_type(call, call->Value(&_gvn));
2146 _gvn.set_type(ret, ret->Value(&_gvn));
2147 // Don't add store to buffer call if we are strength reducing
2148 if (!C->strength_reduction()) {
2149 ideal.if_then(ret, BoolTest::eq, ideal.makecon(TypePtr::NULL_PTR)); {
2150 // Return value is null
2151 ideal.set(res, makecon(TypePtr::NULL_PTR));
2152 } ideal.else_(); {
2153 // Return value is non-null
2154 sync_kit(ideal);
2155
2156 Node* store_to_buf_call = make_runtime_call(RC_NO_LEAF | RC_NO_IO,
2157 OptoRuntime::store_inline_type_fields_Type(),
2158 StubRoutines::store_inline_type_fields_to_buf(),
2159 nullptr, TypePtr::BOTTOM, ret);
2160
2161 // We don't know how many values are returned. This assumes the
2162 // worst case, that all available registers are used.
2163 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2164 if (domain->field_at(i) == Type::HALF) {
2165 store_to_buf_call->init_req(i, top());
2166 continue;
2167 }
2168 Node* proj =_gvn.transform(new ProjNode(call, i));
2169 store_to_buf_call->init_req(i, proj);
2170 }
2171 make_slow_call_ex(store_to_buf_call, env()->Throwable_klass(), false);
2172
2173 Node* buf = _gvn.transform(new ProjNode(store_to_buf_call, TypeFunc::Parms));
2174 const Type* buf_type = TypeOopPtr::make_from_klass(t->as_klass())->join_speculative(TypePtr::NOTNULL);
2175 buf = _gvn.transform(new CheckCastPPNode(control(), buf, buf_type));
2176
2177 ideal.set(res, buf);
2178 ideal.sync_kit(this);
2179 } ideal.end_if();
2180 } else {
2181 for (uint i = TypeFunc::Parms+1; i < domain->cnt(); i++) {
2182 // Will be rewired later in replace_call().
2183 _gvn.transform(new ProjNode(call, i));
2184 }
2185 ideal.set(res, ret);
2186 }
2187 sync_kit(ideal);
2188 ret = _gvn.transform(ideal.value(res));
2189 } else if (!call->method()->return_value_is_larval() && _gvn.type(ret)->is_inlinetypeptr()) {
2190 // In Parse::do_call we call make_from_oop on the final result of the call, but this could be the
2191 // result of merging several call paths. If one of them is made of an actual call node that
2192 // returns an oop, we need to call make_from_oop here as well because we want InlineType
2193 // nodes on every path to avoid merging an unallocated InlineType node path with an oop path.
2194 ret = InlineTypeNode::make_from_oop(this, ret, _gvn.type(ret)->inline_klass());
2195 }
2196 }
2197
2198 return ret;
2199 }
2200
2201 //--------------------set_predefined_input_for_runtime_call--------------------
2202 // Reading and setting the memory state is way conservative here.
2203 // The real problem is that I am not doing real Type analysis on memory,
2204 // so I cannot distinguish card mark stores from other stores. Across a GC
2205 // point the Store Barrier and the card mark memory has to agree. I cannot
2206 // have a card mark store and its barrier split across the GC point from
2207 // either above or below. Here I get that to happen by reading ALL of memory.
2208 // A better answer would be to separate out card marks from other memory.
2209 // For now, return the input memory state, so that it can be reused
2210 // after the call, if this call has restricted memory effects.
2211 Node* GraphKit::set_predefined_input_for_runtime_call(SafePointNode* call, Node* narrow_mem) {
2212 // Set fixed predefined input arguments
2213 call->init_req(TypeFunc::Control, control());
2214 call->init_req(TypeFunc::I_O, top()); // does no i/o
2215 call->init_req(TypeFunc::ReturnAdr, top());
2216 if (call->is_CallLeafPure()) {
2217 call->init_req(TypeFunc::Memory, top());
2279 if (use->is_MergeMem()) {
2280 wl.push(use);
2281 }
2282 }
2283 }
2284
2285 // Replace the call with the current state of the kit.
2286 void GraphKit::replace_call(CallNode* call, Node* result, bool do_replaced_nodes, bool do_asserts) {
2287 JVMState* ejvms = nullptr;
2288 if (has_exceptions()) {
2289 ejvms = transfer_exceptions_into_jvms();
2290 }
2291
2292 ReplacedNodes replaced_nodes = map()->replaced_nodes();
2293 ReplacedNodes replaced_nodes_exception;
2294 Node* ex_ctl = top();
2295
2296 SafePointNode* final_state = stop();
2297
2298 // Find all the needed outputs of this call
2299 CallProjections* callprojs = call->extract_projections(true, do_asserts);
2300
2301 Unique_Node_List wl;
2302 Node* init_mem = call->in(TypeFunc::Memory);
2303 Node* final_mem = final_state->in(TypeFunc::Memory);
2304 Node* final_ctl = final_state->in(TypeFunc::Control);
2305 Node* final_io = final_state->in(TypeFunc::I_O);
2306
2307 // Replace all the old call edges with the edges from the inlining result
2308 if (callprojs->fallthrough_catchproj != nullptr) {
2309 C->gvn_replace_by(callprojs->fallthrough_catchproj, final_ctl);
2310 }
2311 if (callprojs->fallthrough_memproj != nullptr) {
2312 if (final_mem->is_MergeMem()) {
2313 // Parser's exits MergeMem was not transformed but may be optimized
2314 final_mem = _gvn.transform(final_mem);
2315 }
2316 C->gvn_replace_by(callprojs->fallthrough_memproj, final_mem);
2317 add_mergemem_users_to_worklist(wl, final_mem);
2318 }
2319 if (callprojs->fallthrough_ioproj != nullptr) {
2320 C->gvn_replace_by(callprojs->fallthrough_ioproj, final_io);
2321 }
2322
2323 // Replace the result with the new result if it exists and is used
2324 if (callprojs->resproj[0] != nullptr && result != nullptr) {
2325 // If the inlined code is dead, the result projections for an inline type returned as
2326 // fields have not been replaced. They will go away once the call is replaced by TOP below.
2327 assert(callprojs->nb_resproj == 1 || (call->tf()->returns_inline_type_as_fields() && stopped()) ||
2328 (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()),
2329 "unexpected number of results");
2330 // If we are doing strength reduction and the return type is not loaded we
2331 // need to rewire all projections since store_inline_type_fields_to_buf is already present
2332 if (C->strength_reduction() && InlineTypeReturnedAsFields && !call->as_CallJava()->method()->return_type()->is_loaded()) {
2333 const TypeTuple* domain = OptoRuntime::store_inline_type_fields_Type()->domain_cc();
2334 for (uint i = TypeFunc::Parms; i < domain->cnt(); i++) {
2335 C->gvn_replace_by(callprojs->resproj[0], final_state->in(i));
2336 }
2337 } else {
2338 C->gvn_replace_by(callprojs->resproj[0], result);
2339 }
2340 }
2341
2342 if (ejvms == nullptr) {
2343 // No exception edges to simply kill off those paths
2344 if (callprojs->catchall_catchproj != nullptr) {
2345 C->gvn_replace_by(callprojs->catchall_catchproj, C->top());
2346 }
2347 if (callprojs->catchall_memproj != nullptr) {
2348 C->gvn_replace_by(callprojs->catchall_memproj, C->top());
2349 }
2350 if (callprojs->catchall_ioproj != nullptr) {
2351 C->gvn_replace_by(callprojs->catchall_ioproj, C->top());
2352 }
2353 // Replace the old exception object with top
2354 if (callprojs->exobj != nullptr) {
2355 C->gvn_replace_by(callprojs->exobj, C->top());
2356 }
2357 } else {
2358 GraphKit ekit(ejvms);
2359
2360 // Load my combined exception state into the kit, with all phis transformed:
2361 SafePointNode* ex_map = ekit.combine_and_pop_all_exception_states();
2362 replaced_nodes_exception = ex_map->replaced_nodes();
2363
2364 Node* ex_oop = ekit.use_exception_state(ex_map);
2365
2366 if (callprojs->catchall_catchproj != nullptr) {
2367 C->gvn_replace_by(callprojs->catchall_catchproj, ekit.control());
2368 ex_ctl = ekit.control();
2369 }
2370 if (callprojs->catchall_memproj != nullptr) {
2371 Node* ex_mem = ekit.reset_memory();
2372 C->gvn_replace_by(callprojs->catchall_memproj, ex_mem);
2373 add_mergemem_users_to_worklist(wl, ex_mem);
2374 }
2375 if (callprojs->catchall_ioproj != nullptr) {
2376 C->gvn_replace_by(callprojs->catchall_ioproj, ekit.i_o());
2377 }
2378
2379 // Replace the old exception object with the newly created one
2380 if (callprojs->exobj != nullptr) {
2381 C->gvn_replace_by(callprojs->exobj, ex_oop);
2382 }
2383 }
2384
2385 // Disconnect the call from the graph
2386 call->disconnect_inputs(C);
2387 C->gvn_replace_by(call, C->top());
2388
2389 // Clean up any MergeMems that feed other MergeMems since the
2390 // optimizer doesn't like that.
2391 while (wl.size() > 0) {
2392 _gvn.transform(wl.pop());
2393 }
2394
2395 if (callprojs->fallthrough_catchproj != nullptr && !final_ctl->is_top() && do_replaced_nodes) {
2396 replaced_nodes.apply(C, final_ctl);
2397 }
2398 if (!ex_ctl->is_top() && do_replaced_nodes) {
2399 replaced_nodes_exception.apply(C, ex_ctl);
2400 }
2401 }
2402
2403
2404 //------------------------------increment_counter------------------------------
2405 // for statistics: increment a VM counter by 1
2406
2407 void GraphKit::increment_counter(address counter_addr) {
2408 Node* adr1 = makecon(TypeRawPtr::make(counter_addr));
2409 increment_counter(adr1);
2410 }
2411
2412 void GraphKit::increment_counter(Node* counter_addr) {
2413 Node* ctrl = control();
2414 Node* cnt = make_load(ctrl, counter_addr, TypeLong::LONG, T_LONG, MemNode::unordered);
2415 Node* incr = _gvn.transform(new AddLNode(cnt, _gvn.longcon(1)));
2416 store_to_memory(ctrl, counter_addr, incr, T_LONG, MemNode::unordered);
2417 }
2418
2419 void GraphKit::halt(Node* ctrl, Node* frameptr, const char* reason, bool generate_code_in_product) {
2420 Node* halt = new HaltNode(ctrl, frameptr, reason
2421 PRODUCT_ONLY(COMMA generate_code_in_product));
2422 halt = _gvn.transform(halt);
2423 root()->add_req(halt);
2424 if (_gvn.is_IterGVN() != nullptr) {
2425 record_for_igvn(root());
2426 }
2427 }
2428
2429 //------------------------------uncommon_trap----------------------------------
2430 // Bail out to the interpreter in mid-method. Implemented by calling the
2431 // uncommon_trap blob. This helper function inserts a runtime call with the
2432 // right debug info.
2433 Node* GraphKit::uncommon_trap(int trap_request,
2434 ciKlass* klass, const char* comment,
2435 bool must_throw,
2436 bool keep_exact_action) {
2437 if (failing_internal()) {
2438 stop();
2439 }
2440 if (stopped()) return nullptr; // trap reachable?
2441
2442 // Note: If ProfileTraps is true, and if a deopt. actually
2443 // occurs here, the runtime will make sure an MDO exists. There is
2444 // no need to call method()->ensure_method_data() at this point.
2445
2446 // Set the stack pointer to the right value for reexecution:
2588 *
2589 * @param n node that the type applies to
2590 * @param exact_kls type from profiling
2591 * @param maybe_null did profiling see null?
2592 *
2593 * @return node with improved type
2594 */
2595 Node* GraphKit::record_profile_for_speculation(Node* n, ciKlass* exact_kls, ProfilePtrKind ptr_kind) {
2596 const Type* current_type = _gvn.type(n);
2597 assert(UseTypeSpeculation, "type speculation must be on");
2598
2599 const TypePtr* speculative = current_type->speculative();
2600
2601 // Should the klass from the profile be recorded in the speculative type?
2602 if (current_type->would_improve_type(exact_kls, jvms()->depth())) {
2603 const TypeKlassPtr* tklass = TypeKlassPtr::make(exact_kls, Type::trust_interfaces);
2604 const TypeOopPtr* xtype = tklass->as_instance_type();
2605 assert(xtype->klass_is_exact(), "Should be exact");
2606 // Any reason to believe n is not null (from this profiling or a previous one)?
2607 assert(ptr_kind != ProfileAlwaysNull, "impossible here");
2608 const TypePtr* ptr = (ptr_kind != ProfileNeverNull && current_type->speculative_maybe_null()) ? TypePtr::BOTTOM : TypePtr::NOTNULL;
2609 // record the new speculative type's depth
2610 speculative = xtype->cast_to_ptr_type(ptr->ptr())->is_ptr();
2611 speculative = speculative->with_inline_depth(jvms()->depth());
2612 } else if (current_type->would_improve_ptr(ptr_kind)) {
2613 // Profiling report that null was never seen so we can change the
2614 // speculative type to non null ptr.
2615 if (ptr_kind == ProfileAlwaysNull) {
2616 speculative = TypePtr::NULL_PTR;
2617 } else {
2618 assert(ptr_kind == ProfileNeverNull, "nothing else is an improvement");
2619 const TypePtr* ptr = TypePtr::NOTNULL;
2620 if (speculative != nullptr) {
2621 speculative = speculative->cast_to_ptr_type(ptr->ptr())->is_ptr();
2622 } else {
2623 speculative = ptr;
2624 }
2625 }
2626 }
2627
2628 if (speculative != current_type->speculative()) {
2629 // Build a type with a speculative type (what we think we know
2630 // about the type but will need a guard when we use it)
2631 const TypeOopPtr* spec_type = TypeOopPtr::make(TypePtr::BotPTR, Type::Offset::bottom, TypeOopPtr::InstanceBot, speculative);
2632 // We're changing the type, we need a new CheckCast node to carry
2633 // the new type. The new type depends on the control: what
2634 // profiling tells us is only valid from here as far as we can
2635 // tell.
2636 Node* cast = new CheckCastPPNode(control(), n, current_type->remove_speculative()->join_speculative(spec_type));
2637 cast = _gvn.transform(cast);
2638 replace_in_map(n, cast);
2639 n = cast;
2640 }
2641
2642 return n;
2643 }
2644
2645 /**
2646 * Record profiling data from receiver profiling at an invoke with the
2647 * type system so that it can propagate it (speculation)
2648 *
2649 * @param n receiver node
2650 *
2651 * @return node with improved type
2652 */
2653 Node* GraphKit::record_profiled_receiver_for_speculation(Node* n) {
2654 if (!UseTypeSpeculation) {
2655 return n;
2656 }
2657 ciKlass* exact_kls = profile_has_unique_klass();
2658 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2659 if ((java_bc() == Bytecodes::_checkcast ||
2660 java_bc() == Bytecodes::_instanceof ||
2661 java_bc() == Bytecodes::_aastore) &&
2662 method()->method_data()->is_mature()) {
2663 ciProfileData* data = method()->method_data()->bci_to_data(bci());
2664 if (data != nullptr) {
2665 if (java_bc() == Bytecodes::_aastore) {
2666 ciKlass* array_type = nullptr;
2667 ciKlass* element_type = nullptr;
2668 ProfilePtrKind element_ptr = ProfileMaybeNull;
2669 bool flat_array = true;
2670 bool null_free_array = true;
2671 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
2672 exact_kls = element_type;
2673 ptr_kind = element_ptr;
2674 } else {
2675 if (!data->as_BitData()->null_seen()) {
2676 ptr_kind = ProfileNeverNull;
2677 } else {
2678 if (TypeProfileCasts) {
2679 assert(data->is_ReceiverTypeData(), "bad profile data type");
2680 ciReceiverTypeData* call = (ciReceiverTypeData*)data->as_ReceiverTypeData();
2681 uint i = 0;
2682 for (; i < call->row_limit(); i++) {
2683 ciKlass* receiver = call->receiver(i);
2684 if (receiver != nullptr) {
2685 break;
2686 }
2687 }
2688 ptr_kind = (i == call->row_limit()) ? ProfileAlwaysNull : ProfileMaybeNull;
2689 }
2690 }
2691 }
2692 }
2693 }
2694 return record_profile_for_speculation(n, exact_kls, ptr_kind);
2695 }
2696
2697 /**
2698 * Record profiling data from argument profiling at an invoke with the
2699 * type system so that it can propagate it (speculation)
2700 *
2701 * @param dest_method target method for the call
2702 * @param bc what invoke bytecode is this?
2703 */
2704 void GraphKit::record_profiled_arguments_for_speculation(ciMethod* dest_method, Bytecodes::Code bc) {
2705 if (!UseTypeSpeculation) {
2706 return;
2707 }
2708 const TypeFunc* tf = TypeFunc::make(dest_method);
2709 int nargs = tf->domain_sig()->cnt() - TypeFunc::Parms;
2710 int skip = Bytecodes::has_receiver(bc) ? 1 : 0;
2711 for (int j = skip, i = 0; j < nargs && i < TypeProfileArgsLimit; j++) {
2712 const Type *targ = tf->domain_sig()->field_at(j + TypeFunc::Parms);
2713 if (is_reference_type(targ->basic_type())) {
2714 ProfilePtrKind ptr_kind = ProfileMaybeNull;
2715 ciKlass* better_type = nullptr;
2716 if (method()->argument_profiled_type(bci(), i, better_type, ptr_kind)) {
2717 record_profile_for_speculation(argument(j), better_type, ptr_kind);
2718 }
2719 i++;
2720 }
2721 }
2722 }
2723
2724 /**
2725 * Record profiling data from parameter profiling at an invoke with
2726 * the type system so that it can propagate it (speculation)
2727 */
2728 void GraphKit::record_profiled_parameters_for_speculation() {
2729 if (!UseTypeSpeculation) {
2730 return;
2731 }
2732 for (int i = 0, j = 0; i < method()->arg_size() ; i++) {
2852 // The first null ends the list.
2853 Node* parm0, Node* parm1,
2854 Node* parm2, Node* parm3,
2855 Node* parm4, Node* parm5,
2856 Node* parm6, Node* parm7) {
2857 assert(call_addr != nullptr, "must not call null targets");
2858
2859 // Slow-path call
2860 bool is_leaf = !(flags & RC_NO_LEAF);
2861 bool has_io = (!is_leaf && !(flags & RC_NO_IO));
2862 if (call_name == nullptr) {
2863 assert(!is_leaf, "must supply name for leaf");
2864 call_name = OptoRuntime::stub_name(call_addr);
2865 }
2866 CallNode* call;
2867 if (!is_leaf) {
2868 call = new CallStaticJavaNode(call_type, call_addr, call_name, adr_type);
2869 } else if (flags & RC_NO_FP) {
2870 call = new CallLeafNoFPNode(call_type, call_addr, call_name, adr_type);
2871 } else if (flags & RC_VECTOR){
2872 uint num_bits = call_type->range_sig()->field_at(TypeFunc::Parms)->is_vect()->length_in_bytes() * BitsPerByte;
2873 call = new CallLeafVectorNode(call_type, call_addr, call_name, adr_type, num_bits);
2874 } else if (flags & RC_PURE) {
2875 assert(adr_type == nullptr, "pure call does not touch memory");
2876 call = new CallLeafPureNode(call_type, call_addr, call_name);
2877 } else {
2878 call = new CallLeafNode(call_type, call_addr, call_name, adr_type);
2879 }
2880
2881 // The following is similar to set_edges_for_java_call,
2882 // except that the memory effects of the call are restricted to AliasIdxRaw.
2883
2884 // Slow path call has no side-effects, uses few values
2885 bool wide_in = !(flags & RC_NARROW_MEM);
2886 bool wide_out = (C->get_alias_index(adr_type) == Compile::AliasIdxBot);
2887
2888 Node* prev_mem = nullptr;
2889 if (wide_in) {
2890 prev_mem = set_predefined_input_for_runtime_call(call);
2891 } else {
2892 assert(!wide_out, "narrow in => narrow out");
2893 Node* narrow_mem = memory(adr_type);
2894 prev_mem = set_predefined_input_for_runtime_call(call, narrow_mem);
2895 }
2896
2897 // Hook each parm in order. Stop looking at the first null.
2898 if (parm0 != nullptr) { call->init_req(TypeFunc::Parms+0, parm0);
2899 if (parm1 != nullptr) { call->init_req(TypeFunc::Parms+1, parm1);
2900 if (parm2 != nullptr) { call->init_req(TypeFunc::Parms+2, parm2);
2901 if (parm3 != nullptr) { call->init_req(TypeFunc::Parms+3, parm3);
2902 if (parm4 != nullptr) { call->init_req(TypeFunc::Parms+4, parm4);
2903 if (parm5 != nullptr) { call->init_req(TypeFunc::Parms+5, parm5);
2904 if (parm6 != nullptr) { call->init_req(TypeFunc::Parms+6, parm6);
2905 if (parm7 != nullptr) { call->init_req(TypeFunc::Parms+7, parm7);
2906 /* close each nested if ===> */ } } } } } } } }
2907 assert(call->in(call->req()-1) != nullptr || (call->req()-1) > (TypeFunc::Parms+7), "must initialize all parms");
2908
2909 if (!is_leaf) {
2910 // Non-leaves can block and take safepoints:
2911 add_safepoint_edges(call, ((flags & RC_MUST_THROW) != 0));
2912 }
2913 // Non-leaves can throw exceptions:
2914 if (has_io) {
2915 call->set_req(TypeFunc::I_O, i_o());
2916 }
2917
2918 if (flags & RC_UNCOMMON) {
2919 // Set the count to a tiny probability. Cf. Estimate_Block_Frequency.
2920 // (An "if" probability corresponds roughly to an unconditional count.
2921 // Sort of.)
2922 call->set_cnt(PROB_UNLIKELY_MAG(4));
2923 }
2924
2925 Node* c = _gvn.transform(call);
2926 assert(c == call, "cannot disappear");
2927
2935
2936 if (has_io) {
2937 set_i_o(_gvn.transform(new ProjNode(call, TypeFunc::I_O)));
2938 }
2939 return call;
2940
2941 }
2942
2943 // i2b
2944 Node* GraphKit::sign_extend_byte(Node* in) {
2945 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(24)));
2946 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(24)));
2947 }
2948
2949 // i2s
2950 Node* GraphKit::sign_extend_short(Node* in) {
2951 Node* tmp = _gvn.transform(new LShiftINode(in, _gvn.intcon(16)));
2952 return _gvn.transform(new RShiftINode(tmp, _gvn.intcon(16)));
2953 }
2954
2955
2956 //------------------------------merge_memory-----------------------------------
2957 // Merge memory from one path into the current memory state.
2958 void GraphKit::merge_memory(Node* new_mem, Node* region, int new_path) {
2959 for (MergeMemStream mms(merged_memory(), new_mem->as_MergeMem()); mms.next_non_empty2(); ) {
2960 Node* old_slice = mms.force_memory();
2961 Node* new_slice = mms.memory2();
2962 if (old_slice != new_slice) {
2963 PhiNode* phi;
2964 if (old_slice->is_Phi() && old_slice->as_Phi()->region() == region) {
2965 if (mms.is_empty()) {
2966 // clone base memory Phi's inputs for this memory slice
2967 assert(old_slice == mms.base_memory(), "sanity");
2968 phi = PhiNode::make(region, nullptr, Type::MEMORY, mms.adr_type(C));
2969 _gvn.set_type(phi, Type::MEMORY);
2970 for (uint i = 1; i < phi->req(); i++) {
2971 phi->init_req(i, old_slice->in(i));
2972 }
2973 } else {
2974 phi = old_slice->as_Phi(); // Phi was generated already
2975 }
3032 gvn.transform(iff);
3033 if (!bol->is_Con()) gvn.record_for_igvn(iff);
3034 return iff;
3035 }
3036
3037 //-------------------------------gen_subtype_check-----------------------------
3038 // Generate a subtyping check. Takes as input the subtype and supertype.
3039 // Returns 2 values: sets the default control() to the true path and returns
3040 // the false path. Only reads invariant memory; sets no (visible) memory.
3041 // The PartialSubtypeCheckNode sets the hidden 1-word cache in the encoding
3042 // but that's not exposed to the optimizer. This call also doesn't take in an
3043 // Object; if you wish to check an Object you need to load the Object's class
3044 // prior to coming here.
3045 Node* Phase::gen_subtype_check(Node* subklass, Node* superklass, Node** ctrl, Node* mem, PhaseGVN& gvn,
3046 ciMethod* method, int bci) {
3047 Compile* C = gvn.C;
3048 if ((*ctrl)->is_top()) {
3049 return C->top();
3050 }
3051
3052 const TypeKlassPtr* klass_ptr_type = gvn.type(superklass)->is_klassptr();
3053 // For a direct pointer comparison, we need the refined array klass pointer
3054 Node* vm_superklass = superklass;
3055 if (klass_ptr_type->isa_aryklassptr() && klass_ptr_type->klass_is_exact()) {
3056 assert(!klass_ptr_type->is_aryklassptr()->is_refined_type(), "Unexpected refined array klass pointer");
3057 vm_superklass = gvn.makecon(klass_ptr_type->is_aryklassptr()->cast_to_refined_array_klass_ptr());
3058 }
3059
3060 // Fast check for identical types, perhaps identical constants.
3061 // The types can even be identical non-constants, in cases
3062 // involving Array.newInstance, Object.clone, etc.
3063 if (subklass == superklass)
3064 return C->top(); // false path is dead; no test needed.
3065
3066 if (gvn.type(superklass)->singleton()) {
3067 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3068 const TypeKlassPtr* subk = gvn.type(subklass)->is_klassptr();
3069
3070 // In the common case of an exact superklass, try to fold up the
3071 // test before generating code. You may ask, why not just generate
3072 // the code and then let it fold up? The answer is that the generated
3073 // code will necessarily include null checks, which do not always
3074 // completely fold away. If they are also needless, then they turn
3075 // into a performance loss. Example:
3076 // Foo[] fa = blah(); Foo x = fa[0]; fa[1] = x;
3077 // Here, the type of 'fa' is often exact, so the store check
3078 // of fa[1]=x will fold up, without testing the nullness of x.
3079 //
3080 // At macro expansion, we would have already folded the SubTypeCheckNode
3081 // being expanded here because we always perform the static sub type
3082 // check in SubTypeCheckNode::sub() regardless of whether
3083 // StressReflectiveCode is set or not. We can therefore skip this
3084 // static check when StressReflectiveCode is on.
3085 switch (C->static_subtype_check(superk, subk)) {
3086 case Compile::SSC_always_false:
3087 {
3088 Node* always_fail = *ctrl;
3089 *ctrl = gvn.C->top();
3090 return always_fail;
3091 }
3092 case Compile::SSC_always_true:
3093 return C->top();
3094 case Compile::SSC_easy_test:
3095 {
3096 // Just do a direct pointer compare and be done.
3097 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_STATIC_FREQUENT, gvn, T_ADDRESS);
3098 *ctrl = gvn.transform(new IfTrueNode(iff));
3099 return gvn.transform(new IfFalseNode(iff));
3100 }
3101 case Compile::SSC_full_test:
3102 break;
3103 default:
3104 ShouldNotReachHere();
3105 }
3106 }
3107
3108 // %%% Possible further optimization: Even if the superklass is not exact,
3109 // if the subklass is the unique subtype of the superklass, the check
3110 // will always succeed. We could leave a dependency behind to ensure this.
3111
3112 // First load the super-klass's check-offset
3113 Node* p1 = gvn.transform(AddPNode::make_off_heap(superklass, gvn.MakeConX(in_bytes(Klass::super_check_offset_offset()))));
3114 Node* m = C->immutable_memory();
3115 Node* chk_off = gvn.transform(new LoadINode(nullptr, m, p1, gvn.type(p1)->is_ptr(), TypeInt::INT, MemNode::unordered));
3116 int cacheoff_con = in_bytes(Klass::secondary_super_cache_offset());
3117 const TypeInt* chk_off_t = chk_off->Value(&gvn)->isa_int();
3155 gvn.record_for_igvn(r_ok_subtype);
3156
3157 // If we might perform an expensive check, first try to take advantage of profile data that was attached to the
3158 // SubTypeCheck node
3159 if (might_be_cache && method != nullptr && VM_Version::profile_all_receivers_at_type_check()) {
3160 ciCallProfile profile = method->call_profile_at_bci(bci);
3161 float total_prob = 0;
3162 for (int i = 0; profile.has_receiver(i); ++i) {
3163 float prob = profile.receiver_prob(i);
3164 total_prob += prob;
3165 }
3166 if (total_prob * 100. >= TypeProfileSubTypeCheckCommonThreshold) {
3167 const TypeKlassPtr* superk = gvn.type(superklass)->is_klassptr();
3168 for (int i = 0; profile.has_receiver(i); ++i) {
3169 ciKlass* klass = profile.receiver(i);
3170 const TypeKlassPtr* klass_t = TypeKlassPtr::make(klass);
3171 Compile::SubTypeCheckResult result = C->static_subtype_check(superk, klass_t);
3172 if (result != Compile::SSC_always_true && result != Compile::SSC_always_false) {
3173 continue;
3174 }
3175 if (klass_t->isa_aryklassptr()) {
3176 // For a direct pointer comparison, we need the refined array klass pointer
3177 klass_t = klass_t->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3178 }
3179 float prob = profile.receiver_prob(i);
3180 ConNode* klass_node = gvn.makecon(klass_t);
3181 IfNode* iff = gen_subtype_check_compare(*ctrl, subklass, klass_node, BoolTest::eq, prob, gvn, T_ADDRESS);
3182 Node* iftrue = gvn.transform(new IfTrueNode(iff));
3183
3184 if (result == Compile::SSC_always_true) {
3185 r_ok_subtype->add_req(iftrue);
3186 } else {
3187 assert(result == Compile::SSC_always_false, "");
3188 r_not_subtype->add_req(iftrue);
3189 }
3190 *ctrl = gvn.transform(new IfFalseNode(iff));
3191 }
3192 }
3193 }
3194
3195 // See if we get an immediate positive hit. Happens roughly 83% of the
3196 // time. Test to see if the value loaded just previously from the subklass
3197 // is exactly the superklass.
3198 IfNode *iff1 = gen_subtype_check_compare(*ctrl, superklass, nkls, BoolTest::eq, PROB_LIKELY(0.83f), gvn, T_ADDRESS);
3212 igvn->remove_globally_dead_node(r_not_subtype, PhaseIterGVN::NodeOrigin::Speculative);
3213 }
3214 return not_subtype_ctrl;
3215 }
3216
3217 r_ok_subtype->init_req(1, iftrue1);
3218
3219 // Check for immediate negative hit. Happens roughly 11% of the time (which
3220 // is roughly 63% of the remaining cases). Test to see if the loaded
3221 // check-offset points into the subklass display list or the 1-element
3222 // cache. If it points to the display (and NOT the cache) and the display
3223 // missed then it's not a subtype.
3224 Node *cacheoff = gvn.intcon(cacheoff_con);
3225 IfNode *iff2 = gen_subtype_check_compare(*ctrl, chk_off, cacheoff, BoolTest::ne, PROB_LIKELY(0.63f), gvn, T_INT);
3226 r_not_subtype->init_req(1, gvn.transform(new IfTrueNode (iff2)));
3227 *ctrl = gvn.transform(new IfFalseNode(iff2));
3228
3229 // Check for self. Very rare to get here, but it is taken 1/3 the time.
3230 // No performance impact (too rare) but allows sharing of secondary arrays
3231 // which has some footprint reduction.
3232 IfNode *iff3 = gen_subtype_check_compare(*ctrl, subklass, vm_superklass, BoolTest::eq, PROB_LIKELY(0.36f), gvn, T_ADDRESS);
3233 r_ok_subtype->init_req(2, gvn.transform(new IfTrueNode(iff3)));
3234 *ctrl = gvn.transform(new IfFalseNode(iff3));
3235
3236 // -- Roads not taken here: --
3237 // We could also have chosen to perform the self-check at the beginning
3238 // of this code sequence, as the assembler does. This would not pay off
3239 // the same way, since the optimizer, unlike the assembler, can perform
3240 // static type analysis to fold away many successful self-checks.
3241 // Non-foldable self checks work better here in second position, because
3242 // the initial primary superclass check subsumes a self-check for most
3243 // types. An exception would be a secondary type like array-of-interface,
3244 // which does not appear in its own primary supertype display.
3245 // Finally, we could have chosen to move the self-check into the
3246 // PartialSubtypeCheckNode, and from there out-of-line in a platform
3247 // dependent manner. But it is worthwhile to have the check here,
3248 // where it can be perhaps be optimized. The cost in code space is
3249 // small (register compare, branch).
3250
3251 // Now do a linear scan of the secondary super-klass array. Again, no real
3252 // performance impact (too rare) but it's gotta be done.
3253 // Since the code is rarely used, there is no penalty for moving it
3254 // out of line, and it can only improve I-cache density.
3255 // The decision to inline or out-of-line this final check is platform
3256 // dependent, and is found in the AD file definition of PartialSubtypeCheck.
3257 Node* psc = gvn.transform(
3258 new PartialSubtypeCheckNode(*ctrl, subklass, superklass));
3259
3260 IfNode *iff4 = gen_subtype_check_compare(*ctrl, psc, gvn.zerocon(T_OBJECT), BoolTest::ne, PROB_FAIR, gvn, T_ADDRESS);
3261 r_not_subtype->init_req(2, gvn.transform(new IfTrueNode (iff4)));
3262 r_ok_subtype ->init_req(3, gvn.transform(new IfFalseNode(iff4)));
3263
3264 // Return false path; set default control to true path.
3265 *ctrl = gvn.transform(r_ok_subtype);
3266 return gvn.transform(r_not_subtype);
3267 }
3268
3269 Node* GraphKit::gen_subtype_check(Node* obj_or_subklass, Node* superklass) {
3270 const Type* sub_t = _gvn.type(obj_or_subklass);
3271 if (sub_t->make_oopptr() != nullptr && sub_t->make_oopptr()->is_inlinetypeptr()) {
3272 sub_t = TypeKlassPtr::make(sub_t->inline_klass());
3273 obj_or_subklass = makecon(sub_t);
3274 }
3275 bool expand_subtype_check = C->post_loop_opts_phase(); // macro node expansion is over
3276 if (expand_subtype_check) {
3277 MergeMemNode* mem = merged_memory();
3278 Node* ctrl = control();
3279 Node* subklass = obj_or_subklass;
3280 if (!sub_t->isa_klassptr()) {
3281 subklass = load_object_klass(obj_or_subklass);
3282 }
3283
3284 Node* n = Phase::gen_subtype_check(subklass, superklass, &ctrl, mem, _gvn, method(), bci());
3285 set_control(ctrl);
3286 return n;
3287 }
3288
3289 Node* check = _gvn.transform(new SubTypeCheckNode(C, obj_or_subklass, superklass, method(), bci()));
3290 Node* bol = _gvn.transform(new BoolNode(check, BoolTest::eq));
3291 IfNode* iff = create_and_xform_if(control(), bol, PROB_STATIC_FREQUENT, COUNT_UNKNOWN);
3292 set_control(_gvn.transform(new IfTrueNode(iff)));
3293 return _gvn.transform(new IfFalseNode(iff));
3294 }
3295
3296 // Profile-driven exact type check:
3297 Node* GraphKit::type_check_receiver(Node* receiver, ciKlass* klass,
3298 float prob, Node* *casted_receiver) {
3299 assert(!klass->is_interface(), "no exact type check on interfaces");
3300 Node* fail = top();
3301 const Type* rec_t = _gvn.type(receiver);
3302 if (rec_t->is_inlinetypeptr()) {
3303 if (klass->equals(rec_t->inline_klass())) {
3304 (*casted_receiver) = receiver; // Always passes
3305 } else {
3306 (*casted_receiver) = top(); // Always fails
3307 fail = control();
3308 set_control(top());
3309 }
3310 return fail;
3311 }
3312 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces);
3313 if (tklass->isa_aryklassptr()) {
3314 // For a direct pointer comparison, we need the refined array klass pointer
3315 tklass = tklass->is_aryklassptr()->cast_to_refined_array_klass_ptr();
3316 }
3317 Node* recv_klass = load_object_klass(receiver);
3318 fail = type_check(recv_klass, tklass, prob);
3319
3320 if (!stopped()) {
3321 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3322 const TypeOopPtr* recv_xtype = tklass->as_instance_type();
3323 assert(recv_xtype->klass_is_exact(), "");
3324
3325 if (!receiver_type->higher_equal(recv_xtype)) { // ignore redundant casts
3326 // Subsume downstream occurrences of receiver with a cast to
3327 // recv_xtype, since now we know what the type will be.
3328 Node* cast = new CheckCastPPNode(control(), receiver, recv_xtype);
3329 Node* res = _gvn.transform(cast);
3330 if (recv_xtype->is_inlinetypeptr()) {
3331 assert(!gvn().type(res)->maybe_null(), "receiver should never be null");
3332 res = InlineTypeNode::make_from_oop(this, res, recv_xtype->inline_klass());
3333 }
3334 (*casted_receiver) = res;
3335 assert(!(*casted_receiver)->is_top(), "that path should be unreachable");
3336 // (User must make the replace_in_map call.)
3337 }
3338 }
3339
3340 return fail;
3341 }
3342
3343 Node* GraphKit::type_check(Node* recv_klass, const TypeKlassPtr* tklass,
3344 float prob) {
3345 Node* want_klass = makecon(tklass);
3346 Node* cmp = _gvn.transform(new CmpPNode(recv_klass, want_klass));
3347 Node* bol = _gvn.transform(new BoolNode(cmp, BoolTest::eq));
3348 IfNode* iff = create_and_xform_if(control(), bol, prob, COUNT_UNKNOWN);
3349 set_control(_gvn.transform(new IfTrueNode (iff)));
3350 Node* fail = _gvn.transform(new IfFalseNode(iff));
3351 return fail;
3352 }
3353
3354 //------------------------------subtype_check_receiver-------------------------
3355 Node* GraphKit::subtype_check_receiver(Node* receiver, ciKlass* klass,
3356 Node** casted_receiver) {
3357 const TypeKlassPtr* tklass = TypeKlassPtr::make(klass, Type::trust_interfaces)->try_improve();
3358 Node* want_klass = makecon(tklass);
3359
3360 Node* slow_ctl = gen_subtype_check(receiver, want_klass);
3361
3362 // Ignore interface type information until interface types are properly tracked.
3363 if (!stopped() && !klass->is_interface()) {
3364 const TypeOopPtr* receiver_type = _gvn.type(receiver)->isa_oopptr();
3365 const TypeOopPtr* recv_type = tklass->cast_to_exactness(false)->is_klassptr()->as_instance_type();
3366 if (receiver_type != nullptr && !receiver_type->higher_equal(recv_type)) { // ignore redundant casts
3367 Node* cast = _gvn.transform(new CheckCastPPNode(control(), receiver, recv_type));
3368 if (recv_type->is_inlinetypeptr()) {
3369 cast = InlineTypeNode::make_from_oop(this, cast, recv_type->inline_klass());
3370 }
3371 (*casted_receiver) = cast;
3372 }
3373 }
3374
3375 return slow_ctl;
3376 }
3377
3378 //------------------------------seems_never_null-------------------------------
3379 // Use null_seen information if it is available from the profile.
3380 // If we see an unexpected null at a type check we record it and force a
3381 // recompile; the offending check will be recompiled to handle nulls.
3382 // If we see several offending BCIs, then all checks in the
3383 // method will be recompiled.
3384 bool GraphKit::seems_never_null(Node* obj, ciProfileData* data, bool& speculating) {
3385 speculating = !_gvn.type(obj)->speculative_maybe_null();
3386 Deoptimization::DeoptReason reason = Deoptimization::reason_null_check(speculating);
3387 if (UncommonNullCast // Cutout for this technique
3388 && obj != null() // And not the -Xcomp stupid case?
3389 && !too_many_traps(reason)
3390 ) {
3391 if (speculating) {
3460
3461 //------------------------maybe_cast_profiled_receiver-------------------------
3462 // If the profile has seen exactly one type, narrow to exactly that type.
3463 // Subsequent type checks will always fold up.
3464 Node* GraphKit::maybe_cast_profiled_receiver(Node* not_null_obj,
3465 const TypeKlassPtr* require_klass,
3466 ciKlass* spec_klass,
3467 bool safe_for_replace) {
3468 if (!UseTypeProfile || !TypeProfileCasts) return nullptr;
3469
3470 Deoptimization::DeoptReason reason = Deoptimization::reason_class_check(spec_klass != nullptr);
3471
3472 // Make sure we haven't already deoptimized from this tactic.
3473 if (too_many_traps_or_recompiles(reason))
3474 return nullptr;
3475
3476 // (No, this isn't a call, but it's enough like a virtual call
3477 // to use the same ciMethod accessor to get the profile info...)
3478 // If we have a speculative type use it instead of profiling (which
3479 // may not help us)
3480 ciKlass* exact_kls = spec_klass;
3481 if (exact_kls == nullptr) {
3482 if (java_bc() == Bytecodes::_aastore) {
3483 ciKlass* array_type = nullptr;
3484 ciKlass* element_type = nullptr;
3485 ProfilePtrKind element_ptr = ProfileMaybeNull;
3486 bool flat_array = true;
3487 bool null_free_array = true;
3488 method()->array_access_profiled_type(bci(), array_type, element_type, element_ptr, flat_array, null_free_array);
3489 exact_kls = element_type;
3490 } else {
3491 exact_kls = profile_has_unique_klass();
3492 }
3493 }
3494 if (exact_kls != nullptr) {// no cast failures here
3495 if (require_klass == nullptr ||
3496 C->static_subtype_check(require_klass, TypeKlassPtr::make(exact_kls, Type::trust_interfaces)) == Compile::SSC_always_true) {
3497 // If we narrow the type to match what the type profile sees or
3498 // the speculative type, we can then remove the rest of the
3499 // cast.
3500 // This is a win, even if the exact_kls is very specific,
3501 // because downstream operations, such as method calls,
3502 // will often benefit from the sharper type.
3503 Node* exact_obj = not_null_obj; // will get updated in place...
3504 Node* slow_ctl = type_check_receiver(exact_obj, exact_kls, 1.0,
3505 &exact_obj);
3506 { PreserveJVMState pjvms(this);
3507 set_control(slow_ctl);
3508 uncommon_trap_exact(reason, Deoptimization::Action_maybe_recompile);
3509 }
3510 if (safe_for_replace) {
3511 replace_in_map(not_null_obj, exact_obj);
3512 }
3513 return exact_obj;
3603 // If not_null_obj is dead, only null-path is taken
3604 if (stopped()) { // Doing instance-of on a null?
3605 set_control(null_ctl);
3606 return intcon(0);
3607 }
3608 region->init_req(_null_path, null_ctl);
3609 phi ->init_req(_null_path, intcon(0)); // Set null path value
3610 if (null_ctl == top()) {
3611 // Do this eagerly, so that pattern matches like is_diamond_phi
3612 // will work even during parsing.
3613 assert(_null_path == PATH_LIMIT-1, "delete last");
3614 region->del_req(_null_path);
3615 phi ->del_req(_null_path);
3616 }
3617
3618 // Do we know the type check always succeed?
3619 bool known_statically = false;
3620 if (_gvn.type(superklass)->singleton()) {
3621 const TypeKlassPtr* superk = _gvn.type(superklass)->is_klassptr();
3622 const TypeKlassPtr* subk = _gvn.type(obj)->is_oopptr()->as_klass_type();
3623 if (subk != nullptr && subk->is_loaded()) {
3624 int static_res = C->static_subtype_check(superk, subk);
3625 known_statically = (static_res == Compile::SSC_always_true || static_res == Compile::SSC_always_false);
3626 }
3627 }
3628
3629 if (!known_statically) {
3630 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3631 // We may not have profiling here or it may not help us. If we
3632 // have a speculative type use it to perform an exact cast.
3633 ciKlass* spec_obj_type = obj_type->speculative_type();
3634 if (spec_obj_type != nullptr || (ProfileDynamicTypes && data != nullptr)) {
3635 Node* cast_obj = maybe_cast_profiled_receiver(not_null_obj, nullptr, spec_obj_type, safe_for_replace);
3636 if (stopped()) { // Profile disagrees with this path.
3637 set_control(null_ctl); // Null is the only remaining possibility.
3638 return intcon(0);
3639 }
3640 if (cast_obj != nullptr) {
3641 not_null_obj = cast_obj;
3642 }
3643 }
3659 record_for_igvn(region);
3660
3661 // If we know the type check always succeeds then we don't use the
3662 // profiling data at this bytecode. Don't lose it, feed it to the
3663 // type system as a speculative type.
3664 if (safe_for_replace) {
3665 Node* casted_obj = record_profiled_receiver_for_speculation(obj);
3666 replace_in_map(obj, casted_obj);
3667 }
3668
3669 return _gvn.transform(phi);
3670 }
3671
3672 //-------------------------------gen_checkcast---------------------------------
3673 // Generate a checkcast idiom. Used by both the checkcast bytecode and the
3674 // array store bytecode. Stack must be as-if BEFORE doing the bytecode so the
3675 // uncommon-trap paths work. Adjust stack after this call.
3676 // If failure_control is supplied and not null, it is filled in with
3677 // the control edge for the cast failure. Otherwise, an appropriate
3678 // uncommon trap or exception is thrown.
3679 // If 'new_cast_failure_map' is supplied and is not null, it is set to a newly cloned map
3680 // when the current map for the success path is updated with information only present
3681 // on the success path and not the cast failure path. The newly cloned map should then be
3682 // used to emit the uncommon trap in the caller.
3683 Node* GraphKit::gen_checkcast(Node* obj, Node* superklass, Node** failure_control, SafePointNode** new_cast_failure_map, bool null_free, bool maybe_larval) {
3684 assert(new_cast_failure_map == nullptr || failure_control != nullptr,
3685 "failure_control must be set when new_failure_map is used");
3686 kill_dead_locals(); // Benefit all the uncommon traps
3687 const TypeKlassPtr* klass_ptr_type = _gvn.type(superklass)->is_klassptr();
3688 const Type* obj_type = _gvn.type(obj);
3689
3690 const TypeKlassPtr* improved_klass_ptr_type = klass_ptr_type->try_improve();
3691 const TypeOopPtr* toop = improved_klass_ptr_type->cast_to_exactness(false)->as_instance_type();
3692 bool safe_for_replace = (failure_control == nullptr);
3693 assert(!null_free || toop->can_be_inline_type(), "must be an inline type pointer");
3694
3695 // Fast cutout: Check the case that the cast is vacuously true.
3696 // This detects the common cases where the test will short-circuit
3697 // away completely. We do this before we perform the null check,
3698 // because if the test is going to turn into zero code, we don't
3699 // want a residual null check left around. (Causes a slowdown,
3700 // for example, in some objArray manipulations, such as a[i]=a[j].)
3701 if (improved_klass_ptr_type->singleton()) {
3702 const TypeKlassPtr* kptr = nullptr;
3703 if (obj_type->isa_oop_ptr()) {
3704 kptr = obj_type->is_oopptr()->as_klass_type();
3705 } else if (obj->is_InlineType()) {
3706 ciInlineKlass* vk = obj_type->inline_klass();
3707 kptr = TypeInstKlassPtr::make(TypePtr::NotNull, vk, Type::Offset(0));
3708 }
3709
3710 if (kptr != nullptr) {
3711 switch (C->static_subtype_check(improved_klass_ptr_type, kptr)) {
3712 case Compile::SSC_always_true:
3713 // If we know the type check always succeed then we don't use
3714 // the profiling data at this bytecode. Don't lose it, feed it
3715 // to the type system as a speculative type.
3716 obj = record_profiled_receiver_for_speculation(obj);
3717 if (null_free) {
3718 assert(safe_for_replace, "must be");
3719 obj = null_check(obj);
3720 }
3721 assert(stopped() || !toop->is_inlinetypeptr() || obj->is_InlineType(), "should have been scalarized");
3722 return obj;
3723 case Compile::SSC_always_false:
3724 if (null_free) {
3725 assert(safe_for_replace, "must be");
3726 obj = null_check(obj);
3727 }
3728 // It needs a null check because a null will *pass* the cast check.
3729 if (obj_type->isa_oopptr() != nullptr && !obj_type->is_oopptr()->maybe_null()) {
3730 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3731 Deoptimization::DeoptReason reason = is_aastore ?
3732 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3733 builtin_throw(reason);
3734 return top();
3735 } else if (!too_many_traps_or_recompiles(Deoptimization::Reason_null_assert)) {
3736 return null_assert(obj);
3737 }
3738 break; // Fall through to full check
3739 default:
3740 break;
3741 }
3742 }
3743 }
3744
3745 ciProfileData* data = nullptr;
3746 if (failure_control == nullptr) { // use MDO in regular case only
3747 assert(java_bc() == Bytecodes::_aastore ||
3748 java_bc() == Bytecodes::_checkcast,
3749 "interpreter profiles type checks only for these BCs");
3750 if (method()->method_data()->is_mature()) {
3751 data = method()->method_data()->bci_to_data(bci());
3752 }
3753 }
3754
3755 // Make the merge point
3756 enum { _obj_path = 1, _null_path, PATH_LIMIT };
3757 RegionNode* region = new RegionNode(PATH_LIMIT);
3758 Node* phi = new PhiNode(region, toop);
3759 _gvn.set_type(region, Type::CONTROL);
3760 _gvn.set_type(phi, toop);
3761
3762 C->set_has_split_ifs(true); // Has chance for split-if optimization
3763
3764 // Use null-cast information if it is available
3765 bool speculative_not_null = false;
3766 bool never_see_null = ((failure_control == nullptr) // regular case only
3767 && seems_never_null(obj, data, speculative_not_null));
3768
3769 if (obj->is_InlineType()) {
3770 // Re-execute if buffering during triggers deoptimization
3771 PreserveReexecuteState preexecs(this);
3772 jvms()->set_should_reexecute(true);
3773 obj = obj->as_InlineType()->buffer(this, safe_for_replace);
3774 }
3775
3776 // Null check; get casted pointer; set region slot 3
3777 Node* null_ctl = top();
3778 Node* not_null_obj = nullptr;
3779 if (null_free) {
3780 assert(safe_for_replace, "must be");
3781 not_null_obj = null_check(obj);
3782 } else {
3783 not_null_obj = null_check_oop(obj, &null_ctl, never_see_null, safe_for_replace, speculative_not_null);
3784 }
3785
3786 // If not_null_obj is dead, only null-path is taken
3787 if (stopped()) { // Doing instance-of on a null?
3788 set_control(null_ctl);
3789 if (toop->is_inlinetypeptr()) {
3790 return InlineTypeNode::make_null(_gvn, toop->inline_klass());
3791 }
3792 return null();
3793 }
3794 region->init_req(_null_path, null_ctl);
3795 phi ->init_req(_null_path, null()); // Set null path value
3796 if (null_ctl == top()) {
3797 // Do this eagerly, so that pattern matches like is_diamond_phi
3798 // will work even during parsing.
3799 assert(_null_path == PATH_LIMIT-1, "delete last");
3800 region->del_req(_null_path);
3801 phi ->del_req(_null_path);
3802 }
3803
3804 Node* cast_obj = nullptr;
3805 if (improved_klass_ptr_type->klass_is_exact()) {
3806 // The following optimization tries to statically cast the speculative type of the object
3807 // (for example obtained during profiling) to the type of the superklass and then do a
3808 // dynamic check that the type of the object is what we expect. To work correctly
3809 // for checkcast and aastore the type of superklass should be exact.
3810 const TypeOopPtr* obj_type = _gvn.type(obj)->is_oopptr();
3811 // We may not have profiling here or it may not help us. If we have
3812 // a speculative type use it to perform an exact cast.
3813 ciKlass* spec_obj_type = obj_type->speculative_type();
3814 if (spec_obj_type != nullptr || data != nullptr) {
3815 cast_obj = maybe_cast_profiled_receiver(not_null_obj, improved_klass_ptr_type, spec_obj_type, safe_for_replace);
3816 if (cast_obj != nullptr) {
3817 if (failure_control != nullptr) // failure is now impossible
3818 (*failure_control) = top();
3819 // adjust the type of the phi to the exact klass:
3820 phi->raise_bottom_type(_gvn.type(cast_obj)->meet_speculative(TypePtr::NULL_PTR));
3821 }
3822 }
3823 }
3824
3825 if (cast_obj == nullptr) {
3826 // Generate the subtype check
3827 Node* improved_superklass = superklass;
3828 if (improved_klass_ptr_type != klass_ptr_type && improved_klass_ptr_type->singleton()) {
3829 // Only improve the super class for constants which allows subsequent sub type checks to possibly be commoned up.
3830 // The other non-constant cases cannot be improved with a cast node here since they could be folded to top.
3831 // Additionally, the benefit would only be minor in non-constant cases.
3832 improved_superklass = makecon(improved_klass_ptr_type);
3833 }
3834 Node* not_subtype_ctrl = gen_subtype_check(not_null_obj, improved_superklass);
3835 // Plug in success path into the merge
3836 cast_obj = _gvn.transform(new CheckCastPPNode(control(), not_null_obj, toop));
3837 // Failure path ends in uncommon trap (or may be dead - failure impossible)
3838 if (failure_control == nullptr) {
3839 if (not_subtype_ctrl != top()) { // If failure is possible
3840 PreserveJVMState pjvms(this);
3841 set_control(not_subtype_ctrl);
3842 bool is_aastore = (java_bc() == Bytecodes::_aastore);
3843 Deoptimization::DeoptReason reason = is_aastore ?
3844 Deoptimization::Reason_array_check : Deoptimization::Reason_class_check;
3845 builtin_throw(reason);
3846 }
3847 } else {
3848 (*failure_control) = not_subtype_ctrl;
3849 }
3850 }
3851
3852 region->init_req(_obj_path, control());
3853 phi ->init_req(_obj_path, cast_obj);
3854
3855 // A merge of null or Casted-NotNull obj
3856 Node* res = _gvn.transform(phi);
3857
3858 // Note I do NOT always 'replace_in_map(obj,result)' here.
3859 // if( tk->klass()->can_be_primary_super() )
3860 // This means that if I successfully store an Object into an array-of-String
3861 // I 'forget' that the Object is really now known to be a String. I have to
3862 // do this because we don't have true union types for interfaces - if I store
3863 // a Baz into an array-of-Interface and then tell the optimizer it's an
3864 // Interface, I forget that it's also a Baz and cannot do Baz-like field
3865 // references to it. FIX THIS WHEN UNION TYPES APPEAR!
3866 // replace_in_map( obj, res );
3867
3868 // Return final merged results
3869 set_control( _gvn.transform(region) );
3870 record_for_igvn(region);
3871
3872 bool not_inline = !toop->can_be_inline_type();
3873 bool not_flat_in_array = !UseArrayFlattening || not_inline || (toop->is_inlinetypeptr() && !toop->inline_klass()->maybe_flat_in_array());
3874 if (Arguments::is_valhalla_enabled() && (not_inline || not_flat_in_array)) {
3875 // Check if obj has been loaded from an array
3876 obj = obj->isa_DecodeN() ? obj->in(1) : obj;
3877 Node* array = nullptr;
3878 if (obj->isa_Load()) {
3879 Node* address = obj->in(MemNode::Address);
3880 if (address->isa_AddP()) {
3881 array = address->as_AddP()->in(AddPNode::Base);
3882 }
3883 } else if (obj->is_Phi()) {
3884 Node* region = obj->in(0);
3885 // TODO make this more robust (see JDK-8231346)
3886 if (region->req() == 3 && region->in(2) != nullptr && region->in(2)->in(0) != nullptr) {
3887 IfNode* iff = region->in(2)->in(0)->isa_If();
3888 if (iff != nullptr) {
3889 iff->is_flat_array_check(&_gvn, &array);
3890 }
3891 }
3892 }
3893 if (array != nullptr) {
3894 const TypeAryPtr* ary_t = _gvn.type(array)->isa_aryptr();
3895 if (ary_t != nullptr) {
3896 if (!ary_t->is_not_null_free() && !ary_t->is_null_free() && not_inline) {
3897 // Casting array element to a non-inline-type, mark array as not null-free.
3898 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_null_free()));
3899 if (new_cast_failure_map != nullptr) {
3900 // We want to propagate the improved cast node in the current map. Clone it such that we can still properly
3901 // create the cast failure path in the caller without wrongly making the cast node live there.
3902 *new_cast_failure_map = clone_map();
3903 }
3904 replace_in_map(array, cast);
3905 array = cast;
3906 }
3907 if (!ary_t->is_not_flat() && !ary_t->is_flat() && not_flat_in_array) {
3908 // Casting array element to a non-flat-in-array type, mark array as not flat.
3909 Node* cast = _gvn.transform(new CheckCastPPNode(control(), array, ary_t->cast_to_not_flat()));
3910 if (new_cast_failure_map != nullptr && *new_cast_failure_map == nullptr) {
3911 // Same as above.
3912 *new_cast_failure_map = clone_map();
3913 }
3914 replace_in_map(array, cast);
3915 array = cast;
3916 }
3917 }
3918 }
3919 }
3920
3921 if (!stopped() && !res->is_InlineType()) {
3922 res = record_profiled_receiver_for_speculation(res);
3923 if (toop->is_inlinetypeptr() && !maybe_larval) {
3924 Node* vt = InlineTypeNode::make_from_oop(this, res, toop->inline_klass());
3925 res = vt;
3926 if (safe_for_replace) {
3927 replace_in_map(obj, vt);
3928 replace_in_map(not_null_obj, vt);
3929 replace_in_map(res, vt);
3930 }
3931 }
3932 }
3933 return res;
3934 }
3935
3936 Node* GraphKit::mark_word_test(Node* obj, uintptr_t mask_val, bool eq, bool check_lock) {
3937 // Load markword
3938 Node* mark_adr = basic_plus_adr(obj, oopDesc::mark_offset_in_bytes());
3939 Node* mark = make_load(nullptr, mark_adr, TypeX_X, TypeX_X->basic_type(), MemNode::unordered);
3940 if (check_lock && !UseCompactObjectHeaders) {
3941 // COH: Locking does not override the markword with a tagged pointer. We can directly read from the markword.
3942 // Check if obj is locked
3943 Node* locked_bit = MakeConX(markWord::unlocked_value);
3944 locked_bit = _gvn.transform(new AndXNode(locked_bit, mark));
3945 Node* cmp = _gvn.transform(new CmpXNode(locked_bit, MakeConX(0)));
3946 Node* is_unlocked = _gvn.transform(new BoolNode(cmp, BoolTest::ne));
3947 IfNode* iff = new IfNode(control(), is_unlocked, PROB_MAX, COUNT_UNKNOWN);
3948 _gvn.transform(iff);
3949 Node* locked_region = new RegionNode(3);
3950 Node* mark_phi = new PhiNode(locked_region, TypeX_X);
3951
3952 // Unlocked: Use bits from mark word
3953 locked_region->init_req(1, _gvn.transform(new IfTrueNode(iff)));
3954 mark_phi->init_req(1, mark);
3955
3956 // Locked: Load prototype header from klass
3957 set_control(_gvn.transform(new IfFalseNode(iff)));
3958 // Make loads control dependent to make sure they are only executed if array is locked
3959 Node* klass_adr = basic_plus_adr(obj, oopDesc::klass_offset_in_bytes());
3960 Node* klass = _gvn.transform(LoadKlassNode::make(_gvn, C->immutable_memory(), klass_adr, TypeInstPtr::KLASS, TypeInstKlassPtr::OBJECT));
3961 Node* proto_adr = basic_plus_adr(top(), klass, in_bytes(Klass::prototype_header_offset()));
3962 Node* proto = _gvn.transform(LoadNode::make(_gvn, control(), C->immutable_memory(), proto_adr, proto_adr->bottom_type()->is_ptr(), TypeX_X, TypeX_X->basic_type(), MemNode::unordered));
3963
3964 locked_region->init_req(2, control());
3965 mark_phi->init_req(2, proto);
3966 set_control(_gvn.transform(locked_region));
3967 record_for_igvn(locked_region);
3968
3969 mark = mark_phi;
3970 }
3971
3972 // Now check if mark word bits are set
3973 Node* mask = MakeConX(mask_val);
3974 Node* masked = _gvn.transform(new AndXNode(_gvn.transform(mark), mask));
3975 record_for_igvn(masked); // Give it a chance to be optimized out by IGVN
3976 Node* cmp = _gvn.transform(new CmpXNode(masked, mask));
3977 return _gvn.transform(new BoolNode(cmp, eq ? BoolTest::eq : BoolTest::ne));
3978 }
3979
3980 Node* GraphKit::inline_type_test(Node* obj, bool is_inline) {
3981 return mark_word_test(obj, markWord::inline_type_pattern, is_inline, /* check_lock = */ false);
3982 }
3983
3984 Node* GraphKit::flat_array_test(Node* array_or_klass, bool flat) {
3985 // We can't use immutable memory here because the mark word is mutable.
3986 // PhaseIdealLoop::move_flat_array_check_out_of_loop will make sure the
3987 // check is moved out of loops (mainly to enable loop unswitching).
3988 Node* cmp = _gvn.transform(new FlatArrayCheckNode(C, memory(Compile::AliasIdxRaw), array_or_klass));
3989 record_for_igvn(cmp); // Give it a chance to be optimized out by IGVN
3990 return _gvn.transform(new BoolNode(cmp, flat ? BoolTest::eq : BoolTest::ne));
3991 }
3992
3993 Node* GraphKit::null_free_array_test(Node* array, bool null_free) {
3994 return mark_word_test(array, markWord::null_free_array_bit_in_place, null_free);
3995 }
3996
3997 Node* GraphKit::null_free_atomic_array_test(Node* array, ciInlineKlass* vk) {
3998 assert(vk->has_null_free_atomic_layout() || vk->has_null_free_non_atomic_layout(), "Can't be null-free and flat");
3999
4000 // TODO 8350865 Add a stress flag to always access atomic if layout exists?
4001 if (!vk->has_null_free_non_atomic_layout()) {
4002 return intcon(1); // Always atomic
4003 } else if (!vk->has_null_free_atomic_layout()) {
4004 return intcon(0); // Never atomic
4005 }
4006
4007 Node* array_klass = load_object_klass(array);
4008 int layout_kind_offset = in_bytes(FlatArrayKlass::layout_kind_offset());
4009 Node* layout_kind_addr = basic_plus_adr(top(), array_klass, layout_kind_offset);
4010 Node* layout_kind = make_load(nullptr, layout_kind_addr, TypeInt::INT, T_INT, MemNode::unordered);
4011 Node* cmp = _gvn.transform(new CmpINode(layout_kind, intcon((int)LayoutKind::NULL_FREE_ATOMIC_FLAT)));
4012 return _gvn.transform(new BoolNode(cmp, BoolTest::eq));
4013 }
4014
4015 // Deoptimize if 'ary' is a null-free inline type array and 'val' is null
4016 Node* GraphKit::inline_array_null_guard(Node* ary, Node* val, int nargs, bool safe_for_replace) {
4017 RegionNode* region = new RegionNode(3);
4018 Node* null_ctl = top();
4019 null_check_oop(val, &null_ctl);
4020 if (null_ctl != top()) {
4021 PreserveJVMState pjvms(this);
4022 set_control(null_ctl);
4023 {
4024 // Deoptimize if null-free array
4025 BuildCutout unless(this, null_free_array_test(ary, /* null_free = */ false), PROB_MAX);
4026 inc_sp(nargs);
4027 uncommon_trap(Deoptimization::Reason_null_check,
4028 Deoptimization::Action_none);
4029 }
4030 region->init_req(1, control());
4031 }
4032 region->init_req(2, control());
4033 set_control(_gvn.transform(region));
4034 record_for_igvn(region);
4035 if (_gvn.type(val) == TypePtr::NULL_PTR) {
4036 // Since we were just successfully storing null, the array can't be null free.
4037 const TypeAryPtr* ary_t = _gvn.type(ary)->is_aryptr();
4038 ary_t = ary_t->cast_to_not_null_free();
4039 Node* cast = _gvn.transform(new CheckCastPPNode(control(), ary, ary_t));
4040 if (safe_for_replace) {
4041 replace_in_map(ary, cast);
4042 }
4043 ary = cast;
4044 }
4045 return ary;
4046 }
4047
4048 //------------------------------next_monitor-----------------------------------
4049 // What number should be given to the next monitor?
4050 int GraphKit::next_monitor() {
4051 int current = jvms()->monitor_depth()* C->sync_stack_slots();
4052 int next = current + C->sync_stack_slots();
4053 // Keep the toplevel high water mark current:
4054 if (C->fixed_slots() < next) C->set_fixed_slots(next);
4055 return current;
4056 }
4057
4058 //------------------------------insert_mem_bar---------------------------------
4059 // Memory barrier to avoid floating things around
4060 // The membar serves as a pinch point between both control and all memory slices.
4061 Node* GraphKit::insert_mem_bar(int opcode, Node* precedent) {
4062 MemBarNode* mb = MemBarNode::make(C, opcode, Compile::AliasIdxBot, precedent);
4063 mb->init_req(TypeFunc::Control, control());
4064 mb->init_req(TypeFunc::Memory, reset_memory());
4065 Node* membar = _gvn.transform(mb);
4168 lock->create_lock_counter(map()->jvms());
4169 increment_counter(lock->counter()->addr());
4170 }
4171 #endif
4172
4173 return flock;
4174 }
4175
4176
4177 //------------------------------shared_unlock----------------------------------
4178 // Emit unlocking code.
4179 void GraphKit::shared_unlock(Node* box, Node* obj) {
4180 // bci is either a monitorenter bc or InvocationEntryBci
4181 // %%% SynchronizationEntryBCI is redundant; use InvocationEntryBci in interfaces
4182 assert(SynchronizationEntryBCI == InvocationEntryBci, "");
4183
4184 if (stopped()) { // Dead monitor?
4185 map()->pop_monitor(); // Kill monitor from debug info
4186 return;
4187 }
4188 assert(!obj->is_InlineType(), "should not unlock on inline type");
4189
4190 // Memory barrier to avoid floating things down past the locked region
4191 insert_mem_bar(Op_MemBarReleaseLock);
4192
4193 const TypeFunc *tf = OptoRuntime::complete_monitor_exit_Type();
4194 UnlockNode *unlock = new UnlockNode(C, tf);
4195 #ifdef ASSERT
4196 unlock->set_dbg_jvms(sync_jvms());
4197 #endif
4198 uint raw_idx = Compile::AliasIdxRaw;
4199 unlock->init_req( TypeFunc::Control, control() );
4200 unlock->init_req( TypeFunc::Memory , memory(raw_idx) );
4201 unlock->init_req( TypeFunc::I_O , top() ) ; // does no i/o
4202 unlock->init_req( TypeFunc::FramePtr, frameptr() );
4203 unlock->init_req( TypeFunc::ReturnAdr, top() );
4204
4205 unlock->init_req(TypeFunc::Parms + 0, obj);
4206 unlock->init_req(TypeFunc::Parms + 1, box);
4207 unlock = _gvn.transform(unlock)->as_Unlock();
4208
4209 Node* mem = reset_memory();
4210
4211 // unlock has no side-effects, sets few values
4212 set_predefined_output_for_runtime_call(unlock, mem, TypeRawPtr::BOTTOM);
4213
4214 // Kill monitor from debug info
4215 map()->pop_monitor( );
4216 }
4217
4218 //-------------------------------get_layout_helper-----------------------------
4219 // If the given klass is a constant or known to be an array,
4220 // fetch the constant layout helper value into constant_value
4221 // and return null. Otherwise, load the non-constant
4222 // layout helper value, and return the node which represents it.
4223 // This two-faced routine is useful because allocation sites
4224 // almost always feature constant types.
4225 Node* GraphKit::get_layout_helper(Node* klass_node, jint& constant_value) {
4226 const TypeKlassPtr* klass_t = _gvn.type(klass_node)->isa_klassptr();
4227 if (!StressReflectiveCode && klass_t != nullptr) {
4228 bool xklass = klass_t->klass_is_exact();
4229 bool can_be_flat = false;
4230 const TypeAryPtr* ary_type = klass_t->as_instance_type()->isa_aryptr();
4231 if (UseArrayFlattening && !xklass && ary_type != nullptr) {
4232 // Don't constant fold if the runtime type might be a flat array but the static type is not.
4233 const TypeOopPtr* elem = ary_type->elem()->make_oopptr();
4234 can_be_flat = ary_type->can_be_inline_array() && (!elem->is_inlinetypeptr() || elem->inline_klass()->maybe_flat_in_array());
4235 }
4236 if (!can_be_flat && (xklass || (klass_t->isa_aryklassptr() && klass_t->is_aryklassptr()->elem() != Type::BOTTOM))) {
4237 jint lhelper;
4238 if (klass_t->is_flat()) {
4239 lhelper = ary_type->flat_layout_helper();
4240 } else if (klass_t->isa_aryklassptr()) {
4241 BasicType elem = ary_type->elem()->array_element_basic_type();
4242 if (is_reference_type(elem, true)) {
4243 elem = T_OBJECT;
4244 }
4245 lhelper = Klass::array_layout_helper(elem);
4246 } else {
4247 lhelper = klass_t->is_instklassptr()->exact_klass()->layout_helper();
4248 }
4249 if (lhelper != Klass::_lh_neutral_value) {
4250 constant_value = lhelper;
4251 return (Node*) nullptr;
4252 }
4253 }
4254 }
4255 constant_value = Klass::_lh_neutral_value; // put in a known value
4256 Node* lhp = off_heap_plus_addr(klass_node, in_bytes(Klass::layout_helper_offset()));
4257 return make_load(nullptr, lhp, TypeInt::INT, T_INT, MemNode::unordered);
4258 }
4259
4260 // We just put in an allocate/initialize with a big raw-memory effect.
4261 // Hook selected additional alias categories on the initialization.
4262 static void hook_memory_on_init(GraphKit& kit, int alias_idx,
4263 MergeMemNode* init_in_merge,
4264 Node* init_out_raw) {
4265 DEBUG_ONLY(Node* init_in_raw = init_in_merge->base_memory());
4266 assert(init_in_merge->memory_at(alias_idx) == init_in_raw, "");
4267
4268 Node* prevmem = kit.memory(alias_idx);
4269 init_in_merge->set_memory_at(alias_idx, prevmem);
4270 if (init_out_raw != nullptr) {
4271 kit.set_memory(init_out_raw, alias_idx);
4272 }
4273 }
4274
4275 //---------------------------set_output_for_allocation-------------------------
4276 Node* GraphKit::set_output_for_allocation(AllocateNode* alloc,
4277 const TypeOopPtr* oop_type,
4278 bool deoptimize_on_exception) {
4279 int rawidx = Compile::AliasIdxRaw;
4280 alloc->set_req( TypeFunc::FramePtr, frameptr() );
4281 add_safepoint_edges(alloc);
4282 Node* allocx = _gvn.transform(alloc);
4283 set_control( _gvn.transform(new ProjNode(allocx, TypeFunc::Control) ) );
4284 // create memory projection for i_o
4285 set_memory ( _gvn.transform( new ProjNode(allocx, TypeFunc::Memory, true) ), rawidx );
4286 make_slow_call_ex(allocx, env()->Throwable_klass(), true, deoptimize_on_exception);
4287
4288 // create a memory projection as for the normal control path
4289 Node* malloc = _gvn.transform(new ProjNode(allocx, TypeFunc::Memory));
4290 set_memory(malloc, rawidx);
4291
4292 // a normal slow-call doesn't change i_o, but an allocation does
4293 // we create a separate i_o projection for the normal control path
4294 set_i_o(_gvn.transform( new ProjNode(allocx, TypeFunc::I_O, false) ) );
4295 Node* rawoop = _gvn.transform( new ProjNode(allocx, TypeFunc::Parms) );
4296
4297 // put in an initialization barrier
4298 InitializeNode* init = insert_mem_bar_volatile(Op_Initialize, rawidx,
4299 rawoop)->as_Initialize();
4300 assert(alloc->initialization() == init, "2-way macro link must work");
4301 assert(init ->allocation() == alloc, "2-way macro link must work");
4302 {
4303 // Extract memory strands which may participate in the new object's
4304 // initialization, and source them from the new InitializeNode.
4305 // This will allow us to observe initializations when they occur,
4306 // and link them properly (as a group) to the InitializeNode.
4307 assert(init->in(InitializeNode::Memory) == malloc, "");
4308 MergeMemNode* minit_in = MergeMemNode::make(malloc);
4309 init->set_req(InitializeNode::Memory, minit_in);
4310 record_for_igvn(minit_in); // fold it up later, if possible
4311 _gvn.set_type(minit_in, Type::MEMORY);
4312 Node* minit_out = memory(rawidx);
4313 assert(minit_out->is_Proj() && minit_out->in(0) == init, "");
4314 int mark_idx = C->get_alias_index(oop_type->add_offset(oopDesc::mark_offset_in_bytes()));
4315 // Add an edge in the MergeMem for the header fields so an access to one of those has correct memory state.
4316 // Use one NarrowMemProjNode per slice to properly record the adr type of each slice. The Initialize node will have
4317 // multiple projections as a result.
4318 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(mark_idx))), mark_idx);
4319 int klass_idx = C->get_alias_index(oop_type->add_offset(oopDesc::klass_offset_in_bytes()));
4320 set_memory(_gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(klass_idx))), klass_idx);
4321 if (oop_type->isa_aryptr()) {
4322 // Initially all flat array accesses share a single slice
4323 // but that changes after parsing. Prepare the memory graph so
4324 // it can optimize flat array accesses properly once they
4325 // don't share a single slice.
4326 assert(C->flat_accesses_share_alias(), "should be set at parse time");
4327 const TypePtr* telemref = oop_type->add_offset(Type::OffsetBot);
4328 int elemidx = C->get_alias_index(telemref);
4329 const TypePtr* alias_adr_type = C->get_adr_type(elemidx);
4330 if (alias_adr_type->is_flat()) {
4331 C->set_flat_accesses();
4332 }
4333 hook_memory_on_init(*this, elemidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, alias_adr_type)));
4334 } else if (oop_type->isa_instptr()) {
4335 ciInstanceKlass* ik = oop_type->is_instptr()->instance_klass();
4336 for (int i = 0, len = ik->nof_nonstatic_fields(); i < len; i++) {
4337 ciField* field = ik->nonstatic_field_at(i);
4338 if (field->offset_in_bytes() >= TrackedInitializationLimit * HeapWordSize)
4339 continue; // do not bother to track really large numbers of fields
4340 // Find (or create) the alias category for this field:
4341 int fieldidx = C->alias_type(field)->index();
4342 hook_memory_on_init(*this, fieldidx, minit_in, _gvn.transform(new NarrowMemProjNode(init, C->get_adr_type(fieldidx))));
4343 }
4344 }
4345 }
4346
4347 // Cast raw oop to the real thing...
4348 Node* javaoop = new CheckCastPPNode(control(), rawoop, oop_type);
4349 javaoop = _gvn.transform(javaoop);
4350 C->set_recent_alloc(control(), javaoop);
4351 assert(just_allocated_object(control()) == javaoop, "just allocated");
4352
4353 #ifdef ASSERT
4365 assert(alloc->in(AllocateNode::ALength)->is_top(), "no length, please");
4366 }
4367 }
4368 #endif //ASSERT
4369
4370 return javaoop;
4371 }
4372
4373 //---------------------------new_instance--------------------------------------
4374 // This routine takes a klass_node which may be constant (for a static type)
4375 // or may be non-constant (for reflective code). It will work equally well
4376 // for either, and the graph will fold nicely if the optimizer later reduces
4377 // the type to a constant.
4378 // The optional arguments are for specialized use by intrinsics:
4379 // - If 'extra_slow_test' if not null is an extra condition for the slow-path.
4380 // - If 'return_size_val', report the total object size to the caller.
4381 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4382 Node* GraphKit::new_instance(Node* klass_node,
4383 Node* extra_slow_test,
4384 Node* *return_size_val,
4385 bool deoptimize_on_exception,
4386 InlineTypeNode* inline_type_node) {
4387 // Compute size in doublewords
4388 // The size is always an integral number of doublewords, represented
4389 // as a positive bytewise size stored in the klass's layout_helper.
4390 // The layout_helper also encodes (in a low bit) the need for a slow path.
4391 jint layout_con = Klass::_lh_neutral_value;
4392 Node* layout_val = get_layout_helper(klass_node, layout_con);
4393 bool layout_is_con = (layout_val == nullptr);
4394
4395 if (extra_slow_test == nullptr) extra_slow_test = intcon(0);
4396 // Generate the initial go-slow test. It's either ALWAYS (return a
4397 // Node for 1) or NEVER (return a null) or perhaps (in the reflective
4398 // case) a computed value derived from the layout_helper.
4399 Node* initial_slow_test = nullptr;
4400 if (layout_is_con) {
4401 assert(!StressReflectiveCode, "stress mode does not use these paths");
4402 bool must_go_slow = Klass::layout_helper_needs_slow_path(layout_con);
4403 initial_slow_test = must_go_slow ? intcon(1) : extra_slow_test;
4404 } else { // reflective case
4405 // This reflective path is used by Unsafe.allocateInstance.
4406 // (It may be stress-tested by specifying StressReflectiveCode.)
4407 // Basically, we want to get into the VM is there's an illegal argument.
4408 Node* bit = intcon(Klass::_lh_instance_slow_path_bit);
4409 initial_slow_test = _gvn.transform( new AndINode(layout_val, bit) );
4410 if (extra_slow_test != intcon(0)) {
4411 initial_slow_test = _gvn.transform( new OrINode(initial_slow_test, extra_slow_test) );
4412 }
4413 // (Macro-expander will further convert this to a Bool, if necessary.)
4424
4425 // Clear the low bits to extract layout_helper_size_in_bytes:
4426 assert((int)Klass::_lh_instance_slow_path_bit < BytesPerLong, "clear bit");
4427 Node* mask = MakeConX(~ (intptr_t)right_n_bits(LogBytesPerLong));
4428 size = _gvn.transform( new AndXNode(size, mask) );
4429 }
4430 if (return_size_val != nullptr) {
4431 (*return_size_val) = size;
4432 }
4433
4434 // This is a precise notnull oop of the klass.
4435 // (Actually, it need not be precise if this is a reflective allocation.)
4436 // It's what we cast the result to.
4437 const TypeKlassPtr* tklass = _gvn.type(klass_node)->isa_klassptr();
4438 if (!tklass) tklass = TypeInstKlassPtr::OBJECT;
4439 const TypeOopPtr* oop_type = tklass->as_instance_type();
4440
4441 // Now generate allocation code
4442
4443 // The entire memory state is needed for slow path of the allocation
4444 // since GC and deoptimization can happen.
4445 Node *mem = reset_memory();
4446 set_all_memory(mem); // Create new memory state
4447
4448 AllocateNode* alloc = new AllocateNode(C, AllocateNode::alloc_type(Type::TOP),
4449 control(), mem, i_o(),
4450 size, klass_node,
4451 initial_slow_test, inline_type_node);
4452
4453 return set_output_for_allocation(alloc, oop_type, deoptimize_on_exception);
4454 }
4455
4456 //-------------------------------new_array-------------------------------------
4457 // helper for newarray and anewarray
4458 // The 'length' parameter is (obviously) the length of the array.
4459 // The optional arguments are for specialized use by intrinsics:
4460 // - If 'return_size_val', report the non-padded array size (sum of header size
4461 // and array body) to the caller.
4462 // - deoptimize_on_exception controls how Java exceptions are handled (rethrow vs deoptimize)
4463 Node* GraphKit::new_array(Node* klass_node, // array klass (maybe variable)
4464 Node* length, // number of array elements
4465 int nargs, // number of arguments to push back for uncommon trap
4466 Node* *return_size_val,
4467 bool deoptimize_on_exception,
4468 Node* init_val) {
4469 jint layout_con = Klass::_lh_neutral_value;
4470 Node* layout_val = get_layout_helper(klass_node, layout_con);
4471 bool layout_is_con = (layout_val == nullptr);
4472
4473 if (!layout_is_con && !StressReflectiveCode &&
4474 !too_many_traps(Deoptimization::Reason_class_check)) {
4475 // This is a reflective array creation site.
4476 // Optimistically assume that it is a subtype of Object[],
4477 // so that we can fold up all the address arithmetic.
4478 layout_con = Klass::array_layout_helper(T_OBJECT);
4479 Node* cmp_lh = _gvn.transform( new CmpINode(layout_val, intcon(layout_con)) );
4480 Node* bol_lh = _gvn.transform( new BoolNode(cmp_lh, BoolTest::eq) );
4481 { BuildCutout unless(this, bol_lh, PROB_MAX);
4482 inc_sp(nargs);
4483 uncommon_trap(Deoptimization::Reason_class_check,
4484 Deoptimization::Action_maybe_recompile);
4485 }
4486 layout_val = nullptr;
4487 layout_is_con = true;
4488 }
4489
4490 // Generate the initial go-slow test. Make sure we do not overflow
4491 // if length is huge (near 2Gig) or negative! We do not need
4492 // exact double-words here, just a close approximation of needed
4493 // double-words. We can't add any offset or rounding bits, lest we
4494 // take a size -1 of bytes and make it positive. Use an unsigned
4495 // compare, so negative sizes look hugely positive.
4496 int fast_size_limit = FastAllocateSizeLimit;
4497 if (layout_is_con) {
4498 assert(!StressReflectiveCode, "stress mode does not use these paths");
4499 // Increase the size limit if we have exact knowledge of array type.
4500 int log2_esize = Klass::layout_helper_log2_element_size(layout_con);
4501 fast_size_limit <<= MAX2(LogBytesPerLong - log2_esize, 0);
4502 }
4503
4504 Node* initial_slow_cmp = _gvn.transform( new CmpUNode( length, intcon( fast_size_limit ) ) );
4505 Node* initial_slow_test = _gvn.transform( new BoolNode( initial_slow_cmp, BoolTest::gt ) );
4506
4507 // --- Size Computation ---
4508 // array_size = round_to_heap(array_header + (length << elem_shift));
4509 // where round_to_heap(x) == align_to(x, MinObjAlignmentInBytes)
4510 // and align_to(x, y) == ((x + y-1) & ~(y-1))
4511 // The rounding mask is strength-reduced, if possible.
4512 int round_mask = MinObjAlignmentInBytes - 1;
4513 Node* header_size = nullptr;
4514 // (T_BYTE has the weakest alignment and size restrictions...)
4515 if (layout_is_con) {
4516 int hsize = Klass::layout_helper_header_size(layout_con);
4517 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4518 bool is_flat_array = Klass::layout_helper_is_flatArray(layout_con);
4519 if ((round_mask & ~right_n_bits(eshift)) == 0)
4520 round_mask = 0; // strength-reduce it if it goes away completely
4521 assert(is_flat_array || (hsize & right_n_bits(eshift)) == 0, "hsize is pre-rounded");
4522 int header_size_min = arrayOopDesc::base_offset_in_bytes(T_BYTE);
4523 assert(header_size_min <= hsize, "generic minimum is smallest");
4524 header_size = intcon(hsize);
4525 } else {
4526 Node* hss = intcon(Klass::_lh_header_size_shift);
4527 Node* hsm = intcon(Klass::_lh_header_size_mask);
4528 header_size = _gvn.transform(new URShiftINode(layout_val, hss));
4529 header_size = _gvn.transform(new AndINode(header_size, hsm));
4530 }
4531
4532 Node* elem_shift = nullptr;
4533 if (layout_is_con) {
4534 int eshift = Klass::layout_helper_log2_element_size(layout_con);
4535 if (eshift != 0)
4536 elem_shift = intcon(eshift);
4537 } else {
4538 // There is no need to mask or shift this value.
4539 // The semantics of LShiftINode include an implicit mask to 0x1F.
4540 assert(Klass::_lh_log2_element_size_shift == 0, "use shift in place");
4541 elem_shift = layout_val;
4590 }
4591 Node* non_rounded_size = _gvn.transform(new AddXNode(headerx, abody));
4592
4593 if (return_size_val != nullptr) {
4594 // This is the size
4595 (*return_size_val) = non_rounded_size;
4596 }
4597
4598 Node* size = non_rounded_size;
4599 if (round_mask != 0) {
4600 Node* mask1 = MakeConX(round_mask);
4601 size = _gvn.transform(new AddXNode(size, mask1));
4602 Node* mask2 = MakeConX(~round_mask);
4603 size = _gvn.transform(new AndXNode(size, mask2));
4604 }
4605 // else if round_mask == 0, the size computation is self-rounding
4606
4607 // Now generate allocation code
4608
4609 // The entire memory state is needed for slow path of the allocation
4610 // since GC and deoptimization can happen.
4611 Node *mem = reset_memory();
4612 set_all_memory(mem); // Create new memory state
4613
4614 if (initial_slow_test->is_Bool()) {
4615 // Hide it behind a CMoveI, or else PhaseIdealLoop::split_up will get sick.
4616 initial_slow_test = initial_slow_test->as_Bool()->as_int_value(&_gvn);
4617 }
4618
4619 const TypeKlassPtr* ary_klass = _gvn.type(klass_node)->isa_klassptr();
4620 const TypeOopPtr* ary_type = ary_klass->as_instance_type();
4621
4622 Node* raw_init_value = nullptr;
4623 if (init_val != nullptr) {
4624 // TODO 8350865 Fast non-zero init not implemented yet for flat, null-free arrays
4625 if (ary_type->is_flat()) {
4626 initial_slow_test = intcon(1);
4627 }
4628
4629 if (UseCompressedOops) {
4630 // With compressed oops, the 64-bit init value is built from two 32-bit compressed oops
4631 init_val = _gvn.transform(new EncodePNode(init_val, init_val->bottom_type()->make_narrowoop()));
4632 Node* lower = _gvn.transform(new CastP2XNode(control(), init_val));
4633 Node* upper = _gvn.transform(new LShiftLNode(lower, intcon(32)));
4634 raw_init_value = _gvn.transform(new OrLNode(lower, upper));
4635 } else {
4636 raw_init_value = _gvn.transform(new CastP2XNode(control(), init_val));
4637 }
4638 }
4639
4640 Node* valid_length_test = _gvn.intcon(1);
4641 if (ary_type->isa_aryptr()) {
4642 BasicType bt = ary_type->isa_aryptr()->elem()->array_element_basic_type();
4643 jint max = TypeAryPtr::max_array_length(bt);
4644 Node* valid_length_cmp = _gvn.transform(new CmpUNode(length, intcon(max)));
4645 valid_length_test = _gvn.transform(new BoolNode(valid_length_cmp, BoolTest::le));
4646 }
4647
4648 // Create the AllocateArrayNode and its result projections
4649 AllocateArrayNode* alloc
4650 = new AllocateArrayNode(C, AllocateArrayNode::alloc_type(TypeInt::INT),
4651 control(), mem, i_o(),
4652 size, klass_node,
4653 initial_slow_test,
4654 length, valid_length_test,
4655 init_val, raw_init_value);
4656 // Cast to correct type. Note that the klass_node may be constant or not,
4657 // and in the latter case the actual array type will be inexact also.
4658 // (This happens via a non-constant argument to inline_native_newArray.)
4659 // In any case, the value of klass_node provides the desired array type.
4660 const TypeInt* length_type = _gvn.find_int_type(length);
4661 if (ary_type->isa_aryptr() && length_type != nullptr) {
4662 // Try to get a better type than POS for the size
4663 ary_type = ary_type->is_aryptr()->cast_to_size(length_type);
4664 }
4665
4666 Node* javaoop = set_output_for_allocation(alloc, ary_type, deoptimize_on_exception);
4667
4668 array_ideal_length(alloc, ary_type, true);
4669 return javaoop;
4670 }
4671
4672 // The following "Ideal_foo" functions are placed here because they recognize
4673 // the graph shapes created by the functions immediately above.
4674
4675 //---------------------------Ideal_allocation----------------------------------
4770 void GraphKit::add_parse_predicates(int nargs) {
4771 if (ShortRunningLongLoop) {
4772 // Will narrow the limit down with a cast node. Predicates added later may depend on the cast so should be last when
4773 // walking up from the loop.
4774 add_parse_predicate(Deoptimization::Reason_short_running_long_loop, nargs);
4775 }
4776 if (UseLoopPredicate) {
4777 add_parse_predicate(Deoptimization::Reason_predicate, nargs);
4778 if (UseProfiledLoopPredicate) {
4779 add_parse_predicate(Deoptimization::Reason_profile_predicate, nargs);
4780 }
4781 }
4782 if (UseAutoVectorizationPredicate) {
4783 add_parse_predicate(Deoptimization::Reason_auto_vectorization_check, nargs);
4784 }
4785 // Loop Limit Check Predicate should be near the loop.
4786 add_parse_predicate(Deoptimization::Reason_loop_limit_check, nargs);
4787 }
4788
4789 void GraphKit::sync_kit(IdealKit& ideal) {
4790 reset_memory();
4791 set_all_memory(ideal.merged_memory());
4792 set_i_o(ideal.i_o());
4793 set_control(ideal.ctrl());
4794 }
4795
4796 void GraphKit::final_sync(IdealKit& ideal) {
4797 // Final sync IdealKit and graphKit.
4798 sync_kit(ideal);
4799 }
4800
4801 Node* GraphKit::load_String_length(Node* str, bool set_ctrl) {
4802 Node* len = load_array_length(load_String_value(str, set_ctrl));
4803 Node* coder = load_String_coder(str, set_ctrl);
4804 // Divide length by 2 if coder is UTF16
4805 return _gvn.transform(new RShiftINode(len, coder));
4806 }
4807
4808 Node* GraphKit::load_String_value(Node* str, bool set_ctrl) {
4809 int value_offset = java_lang_String::value_offset();
4810 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4811 false, nullptr, Type::Offset(0));
4812 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4813 const TypeAryPtr* value_type = TypeAryPtr::make(TypePtr::NotNull,
4814 TypeAry::make(TypeInt::BYTE, TypeInt::POS, false, false, true, true, true),
4815 ciTypeArrayKlass::make(T_BYTE), true, Type::Offset(0));
4816 Node* p = basic_plus_adr(str, str, value_offset);
4817 Node* load = access_load_at(str, p, value_field_type, value_type, T_OBJECT,
4818 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4819 return load;
4820 }
4821
4822 Node* GraphKit::load_String_coder(Node* str, bool set_ctrl) {
4823 if (!CompactStrings) {
4824 return intcon(java_lang_String::CODER_UTF16);
4825 }
4826 int coder_offset = java_lang_String::coder_offset();
4827 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4828 false, nullptr, Type::Offset(0));
4829 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4830
4831 Node* p = basic_plus_adr(str, str, coder_offset);
4832 Node* load = access_load_at(str, p, coder_field_type, TypeInt::BYTE, T_BYTE,
4833 IN_HEAP | (set_ctrl ? C2_CONTROL_DEPENDENT_LOAD : 0) | MO_UNORDERED);
4834 return load;
4835 }
4836
4837 void GraphKit::store_String_value(Node* str, Node* value) {
4838 int value_offset = java_lang_String::value_offset();
4839 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4840 false, nullptr, Type::Offset(0));
4841 const TypePtr* value_field_type = string_type->add_offset(value_offset);
4842
4843 access_store_at(str, basic_plus_adr(str, value_offset), value_field_type,
4844 value, TypeAryPtr::BYTES, T_OBJECT, IN_HEAP | MO_UNORDERED);
4845 }
4846
4847 void GraphKit::store_String_coder(Node* str, Node* value) {
4848 int coder_offset = java_lang_String::coder_offset();
4849 const TypeInstPtr* string_type = TypeInstPtr::make(TypePtr::NotNull, C->env()->String_klass(),
4850 false, nullptr, Type::Offset(0));
4851 const TypePtr* coder_field_type = string_type->add_offset(coder_offset);
4852
4853 access_store_at(str, basic_plus_adr(str, coder_offset), coder_field_type,
4854 value, TypeInt::BYTE, T_BYTE, IN_HEAP | MO_UNORDERED);
4855 }
4856
4857 // Capture src and dst memory state with a MergeMemNode
4858 Node* GraphKit::capture_memory(const TypePtr* src_type, const TypePtr* dst_type) {
4859 if (src_type == dst_type) {
4860 // Types are equal, we don't need a MergeMemNode
4861 return memory(src_type);
4862 }
4863 MergeMemNode* merge = MergeMemNode::make(map()->memory());
4864 record_for_igvn(merge); // fold it up later, if possible
4865 int src_idx = C->get_alias_index(src_type);
4866 int dst_idx = C->get_alias_index(dst_type);
4867 merge->set_memory_at(src_idx, memory(src_idx));
4868 merge->set_memory_at(dst_idx, memory(dst_idx));
4869 return merge;
4870 }
4943 i_char->init_req(2, AddI(i_char, intcon(2)));
4944
4945 set_control(IfFalse(iff));
4946 set_memory(st, TypeAryPtr::BYTES);
4947 }
4948
4949 Node* GraphKit::make_constant_from_field(ciField* field, Node* obj) {
4950 if (!field->is_constant()) {
4951 return nullptr; // Field not marked as constant.
4952 }
4953 ciInstance* holder = nullptr;
4954 if (!field->is_static()) {
4955 ciObject* const_oop = obj->bottom_type()->is_oopptr()->const_oop();
4956 if (const_oop != nullptr && const_oop->is_instance()) {
4957 holder = const_oop->as_instance();
4958 }
4959 }
4960 const Type* con_type = Type::make_constant_from_field(field, holder, field->layout_type(),
4961 /*is_unsigned_load=*/false);
4962 if (con_type != nullptr) {
4963 Node* con = makecon(con_type);
4964 if (field->type()->is_inlinetype()) {
4965 con = InlineTypeNode::make_from_oop(this, con, field->type()->as_inline_klass());
4966 } else if (con_type->is_inlinetypeptr()) {
4967 con = InlineTypeNode::make_from_oop(this, con, con_type->inline_klass());
4968 }
4969 return con;
4970 }
4971 return nullptr;
4972 }
4973
4974 Node* GraphKit::maybe_narrow_object_type(Node* obj, ciKlass* type, bool maybe_larval) {
4975 const Type* obj_type = obj->bottom_type();
4976 const TypeOopPtr* sig_type = TypeOopPtr::make_from_klass(type);
4977 if (obj_type->isa_oopptr() && sig_type->is_loaded() && !obj_type->higher_equal(sig_type)) {
4978 const Type* narrow_obj_type = obj_type->filter_speculative(sig_type); // keep speculative part
4979 Node* casted_obj = gvn().transform(new CheckCastPPNode(control(), obj, narrow_obj_type));
4980 obj = casted_obj;
4981 }
4982 if (!maybe_larval && sig_type->is_inlinetypeptr()) {
4983 obj = InlineTypeNode::make_from_oop(this, obj, sig_type->inline_klass());
4984 }
4985 return obj;
4986 }
|